YQ Connector: move tests from yql to ydb (OSS)

Перенос папки с тестами на Коннектор из папки yql в папку ydb (синхронизируется с github).

51 files changed, 23016 insertions, 0 deletions

diff --git a/contrib/python/zstandard/py2/.dist-info/METADATA b/contrib/python/zstandard/py2/.dist-info/METADATA
new file mode 100644
index 00000000000..1c38fe3512d
--- /dev/null
+++ b/contrib/python/zstandard/py2/.dist-info/METADATA
@@ -0,0 +1,1637 @@
+Metadata-Version: 2.1
+Name: zstandard
+Version: 0.14.1
+Summary: Zstandard bindings for Python
+Home-page: https://github.com/indygreg/python-zstandard
+Author: Gregory Szorc
+Author-email: [email protected]
+License: BSD
+Keywords: zstandard zstd compression
+Platform: UNKNOWN
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: BSD License
+Classifier: Programming Language :: C
+Classifier: Programming Language :: Python :: 2.7
+Classifier: Programming Language :: Python :: 3.5
+Classifier: Programming Language :: Python :: 3.6
+Classifier: Programming Language :: Python :: 3.7
+Classifier: Programming Language :: Python :: 3.8
+
+================
+python-zstandard
+================
+
+This project provides Python bindings for interfacing with the
+`Zstandard <http://www.zstd.net>`_ compression library. A C extension
+and CFFI interface are provided.
+
+The primary goal of the project is to provide a rich interface to the
+underlying C API through a Pythonic interface while not sacrificing
+performance. This means exposing most of the features and flexibility
+of the C API while not sacrificing usability or safety that Python provides.
+
+The canonical home for this project lives in a Mercurial repository run by
+the author. For convenience, that repository is frequently synchronized to
+https://github.com/indygreg/python-zstandard.
+
+|  |ci-status|
+
+Requirements
+============
+
+This extension is designed to run with Python 2.7, 3.5, 3.6, 3.7, and 3.8
+on common platforms (Linux, Windows, and OS X). On PyPy (both PyPy2 and PyPy3) we support version 6.0.0 and above. 
+x86 and x86_64 are well-tested on Windows. Only x86_64 is well-tested on Linux and macOS.
+
+Installing
+==========
+
+This package is uploaded to PyPI at https://pypi.python.org/pypi/zstandard.
+So, to install this package::
+
+   $ pip install zstandard
+
+Binary wheels are made available for some platforms. If you need to
+install from a source distribution, all you should need is a working C
+compiler and the Python development headers/libraries. On many Linux
+distributions, you can install a ``python-dev`` or ``python-devel``
+package to provide these dependencies.
+
+Packages are also uploaded to Anaconda Cloud at
+https://anaconda.org/indygreg/zstandard. See that URL for how to install
+this package with ``conda``.
+
+Legacy Format Support
+=====================
+
+To enable legacy zstd format support which is needed to handle files compressed
+with zstd < 1.0 you need to provide an installation option::
+
+   $ pip install zstandard --install-option="--legacy"
+
+and since pip 7.0 it is possible to have the following line in your
+requirements.txt::
+
+   zstandard --install-option="--legacy"
+
+Performance
+===========
+
+zstandard is a highly tunable compression algorithm. In its default settings
+(compression level 3), it will be faster at compression and decompression and
+will have better compression ratios than zlib on most data sets. When tuned
+for speed, it approaches lz4's speed and ratios. When tuned for compression
+ratio, it approaches lzma ratios and compression speed, but decompression
+speed is much faster. See the official zstandard documentation for more.
+
+zstandard and this library support multi-threaded compression. There is a
+mechanism to compress large inputs using multiple threads.
+
+The performance of this library is usually very similar to what the zstandard
+C API can deliver. Overhead in this library is due to general Python overhead
+and can't easily be avoided by *any* zstandard Python binding. This library
+exposes multiple APIs for performing compression and decompression so callers
+can pick an API suitable for their need. Contrast with the compression
+modules in Python's standard library (like ``zlib``), which only offer limited
+mechanisms for performing operations. The API flexibility means consumers can
+choose to use APIs that facilitate zero copying or minimize Python object
+creation and garbage collection overhead.
+
+This library is capable of single-threaded throughputs well over 1 GB/s. For
+exact numbers, measure yourself. The source code repository has a ``bench.py``
+script that can be used to measure things.
+
+API
+===
+
+To interface with Zstandard, simply import the ``zstandard`` module::
+
+   import zstandard
+
+It is a popular convention to alias the module as a different name for
+brevity::
+
+   import zstandard as zstd
+
+This module attempts to import and use either the C extension or CFFI
+implementation. On Python platforms known to support C extensions (like
+CPython), it raises an ImportError if the C extension cannot be imported.
+On Python platforms known to not support C extensions (like PyPy), it only
+attempts to import the CFFI implementation and raises ImportError if that
+can't be done. On other platforms, it first tries to import the C extension
+then falls back to CFFI if that fails and raises ImportError if CFFI fails.
+
+To change the module import behavior, a ``PYTHON_ZSTANDARD_IMPORT_POLICY``
+environment variable can be set. The following values are accepted:
+
+default
+   The behavior described above.
+cffi_fallback
+   Always try to import the C extension then fall back to CFFI if that
+   fails.
+cext
+   Only attempt to import the C extension.
+cffi
+   Only attempt to import the CFFI implementation.
+
+In addition, the ``zstandard`` module exports a ``backend`` attribute
+containing the string name of the backend being used. It will be one
+of ``cext`` or ``cffi`` (for *C extension* and *cffi*, respectively).
+
+The types, functions, and attributes exposed by the ``zstandard`` module
+are documented in the sections below.
+
+.. note::
+
+   The documentation in this section makes references to various zstd
+   concepts and functionality. The source repository contains a
+   ``docs/concepts.rst`` file explaining these in more detail.
+
+ZstdCompressor
+--------------
+
+The ``ZstdCompressor`` class provides an interface for performing
+compression operations. Each instance is essentially a wrapper around a
+``ZSTD_CCtx`` from the C API.
+
+Each instance is associated with parameters that control compression
+behavior. These come from the following named arguments (all optional):
+
+level
+   Integer compression level. Valid values are between 1 and 22.
+dict_data
+   Compression dictionary to use.
+
+   Note: When using dictionary data and ``compress()`` is called multiple
+   times, the ``ZstdCompressionParameters`` derived from an integer
+   compression ``level`` and the first compressed data's size will be reused
+   for all subsequent operations. This may not be desirable if source data
+   size varies significantly.
+compression_params
+   A ``ZstdCompressionParameters`` instance defining compression settings.
+write_checksum
+   Whether a 4 byte checksum should be written with the compressed data.
+   Defaults to False. If True, the decompressor can verify that decompressed
+   data matches the original input data.
+write_content_size
+   Whether the size of the uncompressed data will be written into the
+   header of compressed data. Defaults to True. The data will only be
+   written if the compressor knows the size of the input data. This is
+   often not true for streaming compression.
+write_dict_id
+   Whether to write the dictionary ID into the compressed data.
+   Defaults to True. The dictionary ID is only written if a dictionary
+   is being used.
+threads
+   Enables and sets the number of threads to use for multi-threaded compression
+   operations. Defaults to 0, which means to use single-threaded compression.
+   Negative values will resolve to the number of logical CPUs in the system.
+   Read below for more info on multi-threaded compression. This argument only
+   controls thread count for operations that operate on individual pieces of
+   data. APIs that spawn multiple threads for working on multiple pieces of
+   data have their own ``threads`` argument.
+
+``compression_params`` is mutually exclusive with ``level``, ``write_checksum``,
+``write_content_size``, ``write_dict_id``, and ``threads``.
+
+Unless specified otherwise, assume that no two methods of ``ZstdCompressor``
+instances can be called from multiple Python threads simultaneously. In other
+words, assume instances are not thread safe unless stated otherwise.
+
+Utility Methods
+^^^^^^^^^^^^^^^
+
+``frame_progression()`` returns a 3-tuple containing the number of bytes
+ingested, consumed, and produced by the current compression operation.
+
+``memory_size()`` obtains the memory utilization of the underlying zstd
+compression context, in bytes.::
+
+    cctx = zstd.ZstdCompressor()
+    memory = cctx.memory_size()
+
+Simple API
+^^^^^^^^^^
+
+``compress(data)`` compresses and returns data as a one-shot operation.::
+
+   cctx = zstd.ZstdCompressor()
+   compressed = cctx.compress(b'data to compress')
+
+The ``data`` argument can be any object that implements the *buffer protocol*.
+
+Stream Reader API
+^^^^^^^^^^^^^^^^^
+
+``stream_reader(source)`` can be used to obtain an object conforming to the
+``io.RawIOBase`` interface for reading compressed output as a stream::
+
+   with open(path, 'rb') as fh:
+       cctx = zstd.ZstdCompressor()
+       reader = cctx.stream_reader(fh)
+       while True:
+           chunk = reader.read(16384)
+           if not chunk:
+               break
+
+           # Do something with compressed chunk.
+
+Instances can also be used as context managers::
+
+   with open(path, 'rb') as fh:
+       with cctx.stream_reader(fh) as reader:
+           while True:
+               chunk = reader.read(16384)
+               if not chunk:
+                   break
+
+               # Do something with compressed chunk.
+
+When the context manager exits or ``close()`` is called, the stream is closed,
+underlying resources are released, and future operations against the compression
+stream will fail.
+
+The ``source`` argument to ``stream_reader()`` can be any object with a
+``read(size)`` method or any object implementing the *buffer protocol*.
+
+``stream_reader()`` accepts a ``size`` argument specifying how large the input
+stream is. This is used to adjust compression parameters so they are
+tailored to the source size.::
+
+   with open(path, 'rb') as fh:
+       cctx = zstd.ZstdCompressor()
+       with cctx.stream_reader(fh, size=os.stat(path).st_size) as reader:
+           ...
+
+If the ``source`` is a stream, you can specify how large ``read()`` requests
+to that stream should be via the ``read_size`` argument. It defaults to
+``zstandard.COMPRESSION_RECOMMENDED_INPUT_SIZE``.::
+
+   with open(path, 'rb') as fh:
+       cctx = zstd.ZstdCompressor()
+       # Will perform fh.read(8192) when obtaining data to feed into the
+       # compressor.
+       with cctx.stream_reader(fh, read_size=8192) as reader:
+           ...
+
+The stream returned by ``stream_reader()`` is neither writable nor seekable
+(even if the underlying source is seekable). ``readline()`` and
+``readlines()`` are not implemented because they don't make sense for
+compressed data. ``tell()`` returns the number of compressed bytes
+emitted so far.
+
+Streaming Input API
+^^^^^^^^^^^^^^^^^^^
+
+``stream_writer(fh)`` allows you to *stream* data into a compressor.
+
+Returned instances implement the ``io.RawIOBase`` interface. Only methods
+that involve writing will do useful things.
+
+The argument to ``stream_writer()`` must have a ``write(data)`` method. As
+compressed data is available, ``write()`` will be called with the compressed
+data as its argument. Many common Python types implement ``write()``, including
+open file handles and ``io.BytesIO``.
+
+The ``write(data)`` method is used to feed data into the compressor.
+
+The ``flush([flush_mode=FLUSH_BLOCK])`` method can be called to evict whatever
+data remains within the compressor's internal state into the output object. This
+may result in 0 or more ``write()`` calls to the output object. This method
+accepts an optional ``flush_mode`` argument to control the flushing behavior.
+Its value can be any of the ``FLUSH_*`` constants.
+
+Both ``write()`` and ``flush()`` return the number of bytes written to the
+object's ``write()``. In many cases, small inputs do not accumulate enough
+data to cause a write and ``write()`` will return ``0``.
+
+Calling ``close()`` will mark the stream as closed and subsequent I/O
+operations will raise ``ValueError`` (per the documented behavior of
+``io.RawIOBase``). ``close()`` will also call ``close()`` on the underlying
+stream if such a method exists.
+
+Typically usage is as follows::
+
+   cctx = zstd.ZstdCompressor(level=10)
+   compressor = cctx.stream_writer(fh)
+
+   compressor.write(b'chunk 0\n')
+   compressor.write(b'chunk 1\n')
+   compressor.flush()
+   # Receiver will be able to decode ``chunk 0\nchunk 1\n`` at this point.
+   # Receiver is also expecting more data in the zstd *frame*.
+
+   compressor.write(b'chunk 2\n')
+   compressor.flush(zstd.FLUSH_FRAME)
+   # Receiver will be able to decode ``chunk 0\nchunk 1\nchunk 2``.
+   # Receiver is expecting no more data, as the zstd frame is closed.
+   # Any future calls to ``write()`` at this point will construct a new
+   # zstd frame.
+
+Instances can be used as context managers. Exiting the context manager is
+the equivalent of calling ``close()``, which is equivalent to calling
+``flush(zstd.FLUSH_FRAME)``::
+
+   cctx = zstd.ZstdCompressor(level=10)
+   with cctx.stream_writer(fh) as compressor:
+       compressor.write(b'chunk 0')
+       compressor.write(b'chunk 1')
+       ...
+
+.. important::
+
+   If ``flush(FLUSH_FRAME)`` is not called, emitted data doesn't constitute
+   a full zstd *frame* and consumers of this data may complain about malformed
+   input. It is recommended to use instances as a context manager to ensure
+   *frames* are properly finished.
+
+If the size of the data being fed to this streaming compressor is known,
+you can declare it before compression begins::
+
+   cctx = zstd.ZstdCompressor()
+   with cctx.stream_writer(fh, size=data_len) as compressor:
+       compressor.write(chunk0)
+       compressor.write(chunk1)
+       ...
+
+Declaring the size of the source data allows compression parameters to
+be tuned. And if ``write_content_size`` is used, it also results in the
+content size being written into the frame header of the output data.
+
+The size of chunks being ``write()`` to the destination can be specified::
+
+    cctx = zstd.ZstdCompressor()
+    with cctx.stream_writer(fh, write_size=32768) as compressor:
+        ...
+
+To see how much memory is being used by the streaming compressor::
+
+    cctx = zstd.ZstdCompressor()
+    with cctx.stream_writer(fh) as compressor:
+        ...
+        byte_size = compressor.memory_size()
+
+Thte total number of bytes written so far are exposed via ``tell()``::
+
+    cctx = zstd.ZstdCompressor()
+    with cctx.stream_writer(fh) as compressor:
+        ...
+        total_written = compressor.tell()
+
+``stream_writer()`` accepts a ``write_return_read`` boolean argument to control
+the return value of ``write()``. When ``False`` (the default), ``write()`` returns
+the number of bytes that were ``write()``en to the underlying object. When
+``True``, ``write()`` returns the number of bytes read from the input that
+were subsequently written to the compressor. ``True`` is the *proper* behavior
+for ``write()`` as specified by the ``io.RawIOBase`` interface and will become
+the default value in a future release.
+
+Streaming Output API
+^^^^^^^^^^^^^^^^^^^^
+
+``read_to_iter(reader)`` provides a mechanism to stream data out of a
+compressor as an iterator of data chunks.::
+
+   cctx = zstd.ZstdCompressor()
+   for chunk in cctx.read_to_iter(fh):
+        # Do something with emitted data.
+
+``read_to_iter()`` accepts an object that has a ``read(size)`` method or
+conforms to the buffer protocol.
+
+Uncompressed data is fetched from the source either by calling ``read(size)``
+or by fetching a slice of data from the object directly (in the case where
+the buffer protocol is being used). The returned iterator consists of chunks
+of compressed data.
+
+If reading from the source via ``read()``, ``read()`` will be called until
+it raises or returns an empty bytes (``b''``). It is perfectly valid for
+the source to deliver fewer bytes than were what requested by ``read(size)``.
+
+Like ``stream_writer()``, ``read_to_iter()`` also accepts a ``size`` argument
+declaring the size of the input stream::
+
+    cctx = zstd.ZstdCompressor()
+    for chunk in cctx.read_to_iter(fh, size=some_int):
+        pass
+
+You can also control the size that data is ``read()`` from the source and
+the ideal size of output chunks::
+
+    cctx = zstd.ZstdCompressor()
+    for chunk in cctx.read_to_iter(fh, read_size=16384, write_size=8192):
+        pass
+
+Unlike ``stream_writer()``, ``read_to_iter()`` does not give direct control
+over the sizes of chunks fed into the compressor. Instead, chunk sizes will
+be whatever the object being read from delivers. These will often be of a
+uniform size.
+
+Stream Copying API
+^^^^^^^^^^^^^^^^^^
+
+``copy_stream(ifh, ofh)`` can be used to copy data between 2 streams while
+compressing it.::
+
+   cctx = zstd.ZstdCompressor()
+   cctx.copy_stream(ifh, ofh)
+
+For example, say you wish to compress a file::
+
+   cctx = zstd.ZstdCompressor()
+   with open(input_path, 'rb') as ifh, open(output_path, 'wb') as ofh:
+       cctx.copy_stream(ifh, ofh)
+
+It is also possible to declare the size of the source stream::
+
+   cctx = zstd.ZstdCompressor()
+   cctx.copy_stream(ifh, ofh, size=len_of_input)
+
+You can also specify how large the chunks that are ``read()`` and ``write()``
+from and to the streams::
+
+   cctx = zstd.ZstdCompressor()
+   cctx.copy_stream(ifh, ofh, read_size=32768, write_size=16384)
+
+The stream copier returns a 2-tuple of bytes read and written::
+
+   cctx = zstd.ZstdCompressor()
+   read_count, write_count = cctx.copy_stream(ifh, ofh)
+
+Compressor API
+^^^^^^^^^^^^^^
+
+``compressobj()`` returns an object that exposes ``compress(data)`` and
+``flush()`` methods. Each returns compressed data or an empty bytes.
+
+The purpose of ``compressobj()`` is to provide an API-compatible interface
+with ``zlib.compressobj``, ``bz2.BZ2Compressor``, etc. This allows callers to
+swap in different compressor objects while using the same API.
+
+``flush()`` accepts an optional argument indicating how to end the stream.
+``zstd.COMPRESSOBJ_FLUSH_FINISH`` (the default) ends the compression stream.
+Once this type of flush is performed, ``compress()`` and ``flush()`` can
+no longer be called. This type of flush **must** be called to end the
+compression context. If not called, returned data may be incomplete.
+
+A ``zstd.COMPRESSOBJ_FLUSH_BLOCK`` argument to ``flush()`` will flush a
+zstd block. Flushes of this type can be performed multiple times. The next
+call to ``compress()`` will begin a new zstd block.
+
+Here is how this API should be used::
+
+   cctx = zstd.ZstdCompressor()
+   cobj = cctx.compressobj()
+   data = cobj.compress(b'raw input 0')
+   data = cobj.compress(b'raw input 1')
+   data = cobj.flush()
+
+Or to flush blocks::
+
+   cctx.zstd.ZstdCompressor()
+   cobj = cctx.compressobj()
+   data = cobj.compress(b'chunk in first block')
+   data = cobj.flush(zstd.COMPRESSOBJ_FLUSH_BLOCK)
+   data = cobj.compress(b'chunk in second block')
+   data = cobj.flush()
+
+For best performance results, keep input chunks under 256KB. This avoids
+extra allocations for a large output object.
+
+It is possible to declare the input size of the data that will be fed into
+the compressor::
+
+   cctx = zstd.ZstdCompressor()
+   cobj = cctx.compressobj(size=6)
+   data = cobj.compress(b'foobar')
+   data = cobj.flush()
+
+Chunker API
+^^^^^^^^^^^
+
+``chunker(size=None, chunk_size=COMPRESSION_RECOMMENDED_OUTPUT_SIZE)`` returns
+an object that can be used to iteratively feed chunks of data into a compressor
+and produce output chunks of a uniform size.
+
+The object returned by ``chunker()`` exposes the following methods:
+
+``compress(data)``
+   Feeds new input data into the compressor.
+
+``flush()``
+   Flushes all data currently in the compressor.
+
+``finish()``
+   Signals the end of input data. No new data can be compressed after this
+   method is called.
+
+``compress()``, ``flush()``, and ``finish()`` all return an iterator of
+``bytes`` instances holding compressed data. The iterator may be empty. Callers
+MUST iterate through all elements of the returned iterator before performing
+another operation on the object.
+
+All chunks emitted by ``compress()`` will have a length of ``chunk_size``.
+
+``flush()`` and ``finish()`` may return a final chunk smaller than
+``chunk_size``.
+
+Here is how the API should be used::
+
+   cctx = zstd.ZstdCompressor()
+   chunker = cctx.chunker(chunk_size=32768)
+
+   with open(path, 'rb') as fh:
+       while True:
+           in_chunk = fh.read(32768)
+           if not in_chunk:
+               break
+
+           for out_chunk in chunker.compress(in_chunk):
+               # Do something with output chunk of size 32768.
+
+       for out_chunk in chunker.finish():
+           # Do something with output chunks that finalize the zstd frame.
+
+The ``chunker()`` API is often a better alternative to ``compressobj()``.
+
+``compressobj()`` will emit output data as it is available. This results in a
+*stream* of output chunks of varying sizes. The consistency of the output chunk
+size with ``chunker()`` is more appropriate for many usages, such as sending
+compressed data to a socket.
+
+``compressobj()`` may also perform extra memory reallocations in order to
+dynamically adjust the sizes of the output chunks. Since ``chunker()`` output
+chunks are all the same size (except for flushed or final chunks), there is
+less memory allocation overhead.
+
+Batch Compression API
+^^^^^^^^^^^^^^^^^^^^^
+
+(Experimental. Not yet supported in CFFI bindings.)
+
+``multi_compress_to_buffer(data, [threads=0])`` performs compression of multiple
+inputs as a single operation.
+
+Data to be compressed can be passed as a ``BufferWithSegmentsCollection``, a
+``BufferWithSegments``, or a list containing byte like objects. Each element of
+the container will be compressed individually using the configured parameters
+on the ``ZstdCompressor`` instance.
+
+The ``threads`` argument controls how many threads to use for compression. The
+default is ``0`` which means to use a single thread. Negative values use the
+number of logical CPUs in the machine.
+
+The function returns a ``BufferWithSegmentsCollection``. This type represents
+N discrete memory allocations, eaching holding 1 or more compressed frames.
+
+Output data is written to shared memory buffers. This means that unlike
+regular Python objects, a reference to *any* object within the collection
+keeps the shared buffer and therefore memory backing it alive. This can have
+undesirable effects on process memory usage.
+
+The API and behavior of this function is experimental and will likely change.
+Known deficiencies include:
+
+* If asked to use multiple threads, it will always spawn that many threads,
+  even if the input is too small to use them. It should automatically lower
+  the thread count when the extra threads would just add overhead.
+* The buffer allocation strategy is fixed. There is room to make it dynamic,
+  perhaps even to allow one output buffer per input, facilitating a variation
+  of the API to return a list without the adverse effects of shared memory
+  buffers.
+
+ZstdDecompressor
+----------------
+
+The ``ZstdDecompressor`` class provides an interface for performing
+decompression. It is effectively a wrapper around the ``ZSTD_DCtx`` type from
+the C API.
+
+Each instance is associated with parameters that control decompression. These
+come from the following named arguments (all optional):
+
+dict_data
+   Compression dictionary to use.
+max_window_size
+   Sets an uppet limit on the window size for decompression operations in
+   kibibytes. This setting can be used to prevent large memory allocations
+   for inputs using large compression windows.
+format
+   Set the format of data for the decoder. By default, this is
+   ``zstd.FORMAT_ZSTD1``. It can be set to ``zstd.FORMAT_ZSTD1_MAGICLESS`` to
+   allow decoding frames without the 4 byte magic header. Not all decompression
+   APIs support this mode.
+
+The interface of this class is very similar to ``ZstdCompressor`` (by design).
+
+Unless specified otherwise, assume that no two methods of ``ZstdDecompressor``
+instances can be called from multiple Python threads simultaneously. In other
+words, assume instances are not thread safe unless stated otherwise.
+
+Utility Methods
+^^^^^^^^^^^^^^^
+
+``memory_size()`` obtains the size of the underlying zstd decompression context,
+in bytes.::
+
+    dctx = zstd.ZstdDecompressor()
+    size = dctx.memory_size()
+
+Simple API
+^^^^^^^^^^
+
+``decompress(data)`` can be used to decompress an entire compressed zstd
+frame in a single operation.::
+
+    dctx = zstd.ZstdDecompressor()
+    decompressed = dctx.decompress(data)
+
+By default, ``decompress(data)`` will only work on data written with the content
+size encoded in its header (this is the default behavior of
+``ZstdCompressor().compress()`` but may not be true for streaming compression). If
+compressed data without an embedded content size is seen, ``zstd.ZstdError`` will
+be raised.
+
+If the compressed data doesn't have its content size embedded within it,
+decompression can be attempted by specifying the ``max_output_size``
+argument.::
+
+    dctx = zstd.ZstdDecompressor()
+    uncompressed = dctx.decompress(data, max_output_size=1048576)
+
+Ideally, ``max_output_size`` will be identical to the decompressed output
+size.
+
+If ``max_output_size`` is too small to hold the decompressed data,
+``zstd.ZstdError`` will be raised.
+
+If ``max_output_size`` is larger than the decompressed data, the allocated
+output buffer will be resized to only use the space required.
+
+Please note that an allocation of the requested ``max_output_size`` will be
+performed every time the method is called. Setting to a very large value could
+result in a lot of work for the memory allocator and may result in
+``MemoryError`` being raised if the allocation fails.
+
+.. important::
+
+   If the exact size of decompressed data is unknown (not passed in explicitly
+   and not stored in the zstandard frame), for performance reasons it is
+   encouraged to use a streaming API.
+
+Stream Reader API
+^^^^^^^^^^^^^^^^^
+
+``stream_reader(source)`` can be used to obtain an object conforming to the
+``io.RawIOBase`` interface for reading decompressed output as a stream::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       reader = dctx.stream_reader(fh)
+       while True:
+           chunk = reader.read(16384)
+            if not chunk:
+                break
+
+            # Do something with decompressed chunk.
+
+The stream can also be used as a context manager::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       with dctx.stream_reader(fh) as reader:
+           ...
+
+When used as a context manager, the stream is closed and the underlying
+resources are released when the context manager exits. Future operations against
+the stream will fail.
+
+The ``source`` argument to ``stream_reader()`` can be any object with a
+``read(size)`` method or any object implementing the *buffer protocol*.
+
+If the ``source`` is a stream, you can specify how large ``read()`` requests
+to that stream should be via the ``read_size`` argument. It defaults to
+``zstandard.DECOMPRESSION_RECOMMENDED_INPUT_SIZE``.::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       # Will perform fh.read(8192) when obtaining data for the decompressor.
+       with dctx.stream_reader(fh, read_size=8192) as reader:
+           ...
+
+The stream returned by ``stream_reader()`` is not writable.
+
+The stream returned by ``stream_reader()`` is *partially* seekable.
+Absolute and relative positions (``SEEK_SET`` and ``SEEK_CUR``) forward
+of the current position are allowed. Offsets behind the current read
+position and offsets relative to the end of stream are not allowed and
+will raise ``ValueError`` if attempted.
+
+``tell()`` returns the number of decompressed bytes read so far.
+
+Not all I/O methods are implemented. Notably missing is support for
+``readline()``, ``readlines()``, and linewise iteration support. This is
+because streams operate on binary data - not text data. If you want to
+convert decompressed output to text, you can chain an ``io.TextIOWrapper``
+to the stream::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       stream_reader = dctx.stream_reader(fh)
+       text_stream = io.TextIOWrapper(stream_reader, encoding='utf-8')
+
+       for line in text_stream:
+           ...
+
+The ``read_across_frames`` argument to ``stream_reader()`` controls the
+behavior of read operations when the end of a zstd *frame* is encountered.
+When ``False`` (the default), a read will complete when the end of a
+zstd *frame* is encountered. When ``True``, a read can potentially
+return data spanning multiple zstd *frames*.
+
+Streaming Input API
+^^^^^^^^^^^^^^^^^^^
+
+``stream_writer(fh)`` allows you to *stream* data into a decompressor.
+
+Returned instances implement the ``io.RawIOBase`` interface. Only methods
+that involve writing will do useful things.
+
+The argument to ``stream_writer()`` is typically an object that also implements
+``io.RawIOBase``. But any object with a ``write(data)`` method will work. Many
+common Python types conform to this interface, including open file handles
+and ``io.BytesIO``.
+
+Behavior is similar to ``ZstdCompressor.stream_writer()``: compressed data
+is sent to the decompressor by calling ``write(data)`` and decompressed
+output is written to the underlying stream by calling its ``write(data)``
+method.::
+
+    dctx = zstd.ZstdDecompressor()
+    decompressor = dctx.stream_writer(fh)
+
+    decompressor.write(compressed_data)
+    ...
+
+
+Calls to ``write()`` will return the number of bytes written to the output
+object. Not all inputs will result in bytes being written, so return values
+of ``0`` are possible.
+
+Like the ``stream_writer()`` compressor, instances can be used as context
+managers. However, context managers add no extra special behavior and offer
+little to no benefit to being used.
+
+Calling ``close()`` will mark the stream as closed and subsequent I/O operations
+will raise ``ValueError`` (per the documented behavior of ``io.RawIOBase``).
+``close()`` will also call ``close()`` on the underlying stream if such a
+method exists.
+
+The size of chunks being ``write()`` to the destination can be specified::
+
+    dctx = zstd.ZstdDecompressor()
+    with dctx.stream_writer(fh, write_size=16384) as decompressor:
+        pass
+
+You can see how much memory is being used by the decompressor::
+
+    dctx = zstd.ZstdDecompressor()
+    with dctx.stream_writer(fh) as decompressor:
+        byte_size = decompressor.memory_size()
+
+``stream_writer()`` accepts a ``write_return_read`` boolean argument to control
+the return value of ``write()``. When ``False`` (the default)``, ``write()``
+returns the number of bytes that were ``write()``en to the underlying stream.
+When ``True``, ``write()`` returns the number of bytes read from the input.
+``True`` is the *proper* behavior for ``write()`` as specified by the
+``io.RawIOBase`` interface and will become the default in a future release.
+
+Streaming Output API
+^^^^^^^^^^^^^^^^^^^^
+
+``read_to_iter(fh)`` provides a mechanism to stream decompressed data out of a
+compressed source as an iterator of data chunks.:: 
+
+    dctx = zstd.ZstdDecompressor()
+    for chunk in dctx.read_to_iter(fh):
+        # Do something with original data.
+
+``read_to_iter()`` accepts an object with a ``read(size)`` method that will
+return  compressed bytes or an object conforming to the buffer protocol that
+can expose its data as a contiguous range of bytes.
+
+``read_to_iter()`` returns an iterator whose elements are chunks of the
+decompressed data.
+
+The size of requested ``read()`` from the source can be specified::
+
+    dctx = zstd.ZstdDecompressor()
+    for chunk in dctx.read_to_iter(fh, read_size=16384):
+        pass
+
+It is also possible to skip leading bytes in the input data::
+
+    dctx = zstd.ZstdDecompressor()
+    for chunk in dctx.read_to_iter(fh, skip_bytes=1):
+        pass
+
+.. tip::
+
+   Skipping leading bytes is useful if the source data contains extra
+   *header* data. Traditionally, you would need to create a slice or
+   ``memoryview`` of the data you want to decompress. This would create
+   overhead. It is more efficient to pass the offset into this API.
+
+Similarly to ``ZstdCompressor.read_to_iter()``, the consumer of the iterator
+controls when data is decompressed. If the iterator isn't consumed,
+decompression is put on hold.
+
+When ``read_to_iter()`` is passed an object conforming to the buffer protocol,
+the behavior may seem similar to what occurs when the simple decompression
+API is used. However, this API works when the decompressed size is unknown.
+Furthermore, if feeding large inputs, the decompressor will work in chunks
+instead of performing a single operation.
+
+Stream Copying API
+^^^^^^^^^^^^^^^^^^
+
+``copy_stream(ifh, ofh)`` can be used to copy data across 2 streams while
+performing decompression.::
+
+    dctx = zstd.ZstdDecompressor()
+    dctx.copy_stream(ifh, ofh)
+
+e.g. to decompress a file to another file::
+
+    dctx = zstd.ZstdDecompressor()
+    with open(input_path, 'rb') as ifh, open(output_path, 'wb') as ofh:
+        dctx.copy_stream(ifh, ofh)
+
+The size of chunks being ``read()`` and ``write()`` from and to the streams
+can be specified::
+
+    dctx = zstd.ZstdDecompressor()
+    dctx.copy_stream(ifh, ofh, read_size=8192, write_size=16384)
+
+Decompressor API
+^^^^^^^^^^^^^^^^
+
+``decompressobj()`` returns an object that exposes a ``decompress(data)``
+method. Compressed data chunks are fed into ``decompress(data)`` and
+uncompressed output (or an empty bytes) is returned. Output from subsequent
+calls needs to be concatenated to reassemble the full decompressed byte
+sequence.
+
+The purpose of ``decompressobj()`` is to provide an API-compatible interface
+with ``zlib.decompressobj`` and ``bz2.BZ2Decompressor``. This allows callers
+to swap in different decompressor objects while using the same API.
+
+Each object is single use: once an input frame is decoded, ``decompress()``
+can no longer be called.
+
+Here is how this API should be used::
+
+   dctx = zstd.ZstdDecompressor()
+   dobj = dctx.decompressobj()
+   data = dobj.decompress(compressed_chunk_0)
+   data = dobj.decompress(compressed_chunk_1)
+
+By default, calls to ``decompress()`` write output data in chunks of size
+``DECOMPRESSION_RECOMMENDED_OUTPUT_SIZE``. These chunks are concatenated
+before being returned to the caller. It is possible to define the size of
+these temporary chunks by passing ``write_size`` to ``decompressobj()``::
+
+   dctx = zstd.ZstdDecompressor()
+   dobj = dctx.decompressobj(write_size=1048576)
+
+.. note::
+
+   Because calls to ``decompress()`` may need to perform multiple
+   memory (re)allocations, this streaming decompression API isn't as
+   efficient as other APIs.
+
+For compatibility with the standard library APIs, instances expose a
+``flush([length=None])`` method. This method no-ops and has no meaningful
+side-effects, making it safe to call any time.
+
+Batch Decompression API
+^^^^^^^^^^^^^^^^^^^^^^^
+
+(Experimental. Not yet supported in CFFI bindings.)
+
+``multi_decompress_to_buffer()`` performs decompression of multiple
+frames as a single operation and returns a ``BufferWithSegmentsCollection``
+containing decompressed data for all inputs.
+
+Compressed frames can be passed to the function as a ``BufferWithSegments``,
+a ``BufferWithSegmentsCollection``, or as a list containing objects that
+conform to the buffer protocol. For best performance, pass a
+``BufferWithSegmentsCollection`` or a ``BufferWithSegments``, as
+minimal input validation will be done for that type. If calling from
+Python (as opposed to C), constructing one of these instances may add
+overhead cancelling out the performance overhead of validation for list
+inputs.::
+
+    dctx = zstd.ZstdDecompressor()
+    results = dctx.multi_decompress_to_buffer([b'...', b'...'])
+
+The decompressed size of each frame MUST be discoverable. It can either be
+embedded within the zstd frame (``write_content_size=True`` argument to
+``ZstdCompressor``) or passed in via the ``decompressed_sizes`` argument.
+
+The ``decompressed_sizes`` argument is an object conforming to the buffer
+protocol which holds an array of 64-bit unsigned integers in the machine's
+native format defining the decompressed sizes of each frame. If this argument
+is passed, it avoids having to scan each frame for its decompressed size.
+This frame scanning can add noticeable overhead in some scenarios.::
+
+    frames = [...]
+    sizes = struct.pack('=QQQQ', len0, len1, len2, len3)
+
+    dctx = zstd.ZstdDecompressor()
+    results = dctx.multi_decompress_to_buffer(frames, decompressed_sizes=sizes)
+
+The ``threads`` argument controls the number of threads to use to perform
+decompression operations. The default (``0``) or the value ``1`` means to
+use a single thread. Negative values use the number of logical CPUs in the
+machine.
+
+.. note::
+
+   It is possible to pass a ``mmap.mmap()`` instance into this function by
+   wrapping it with a ``BufferWithSegments`` instance (which will define the
+   offsets of frames within the memory mapped region).
+
+This function is logically equivalent to performing ``dctx.decompress()``
+on each input frame and returning the result.
+
+This function exists to perform decompression on multiple frames as fast
+as possible by having as little overhead as possible. Since decompression is
+performed as a single operation and since the decompressed output is stored in
+a single buffer, extra memory allocations, Python objects, and Python function
+calls are avoided. This is ideal for scenarios where callers know up front that
+they need to access data for multiple frames, such as when  *delta chains* are
+being used.
+
+Currently, the implementation always spawns multiple threads when requested,
+even if the amount of work to do is small. In the future, it will be smarter
+about avoiding threads and their associated overhead when the amount of
+work to do is small.
+
+Prefix Dictionary Chain Decompression
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``decompress_content_dict_chain(frames)`` performs decompression of a list of
+zstd frames produced using chained *prefix* dictionary compression. Such
+a list of frames is produced by compressing discrete inputs where each
+non-initial input is compressed with a *prefix* dictionary consisting of the
+content of the previous input.
+
+For example, say you have the following inputs::
+
+   inputs = [b'input 1', b'input 2', b'input 3']
+
+The zstd frame chain consists of:
+
+1. ``b'input 1'`` compressed in standalone/discrete mode
+2. ``b'input 2'`` compressed using ``b'input 1'`` as a *prefix* dictionary
+3. ``b'input 3'`` compressed using ``b'input 2'`` as a *prefix* dictionary
+
+Each zstd frame **must** have the content size written.
+
+The following Python code can be used to produce a *prefix dictionary chain*::
+
+    def make_chain(inputs):
+        frames = []
+
+        # First frame is compressed in standalone/discrete mode.
+        zctx = zstd.ZstdCompressor()
+        frames.append(zctx.compress(inputs[0]))
+
+        # Subsequent frames use the previous fulltext as a prefix dictionary
+        for i, raw in enumerate(inputs[1:]):
+            dict_data = zstd.ZstdCompressionDict(
+                inputs[i], dict_type=zstd.DICT_TYPE_RAWCONTENT)
+            zctx = zstd.ZstdCompressor(dict_data=dict_data)
+            frames.append(zctx.compress(raw))
+
+        return frames
+
+``decompress_content_dict_chain()`` returns the uncompressed data of the last
+element in the input chain.
+
+
+.. note::
+
+   It is possible to implement *prefix dictionary chain* decompression
+   on top of other APIs. However, this function will likely be faster -
+   especially for long input chains - as it avoids the overhead of instantiating
+   and passing around intermediate objects between C and Python.
+
+Multi-Threaded Compression
+--------------------------
+
+``ZstdCompressor`` accepts a ``threads`` argument that controls the number
+of threads to use for compression. The way this works is that input is split
+into segments and each segment is fed into a worker pool for compression. Once
+a segment is compressed, it is flushed/appended to the output.
+
+.. note::
+
+   These threads are created at the C layer and are not Python threads. So they
+   work outside the GIL. It is therefore possible to CPU saturate multiple cores
+   from Python.
+
+The segment size for multi-threaded compression is chosen from the window size
+of the compressor. This is derived from the ``window_log`` attribute of a
+``ZstdCompressionParameters`` instance. By default, segment sizes are in the 1+MB
+range.
+
+If multi-threaded compression is requested and the input is smaller than the
+configured segment size, only a single compression thread will be used. If the
+input is smaller than the segment size multiplied by the thread pool size or
+if data cannot be delivered to the compressor fast enough, not all requested
+compressor threads may be active simultaneously.
+
+Compared to non-multi-threaded compression, multi-threaded compression has
+higher per-operation overhead. This includes extra memory operations,
+thread creation, lock acquisition, etc.
+
+Due to the nature of multi-threaded compression using *N* compression
+*states*, the output from multi-threaded compression will likely be larger
+than non-multi-threaded compression. The difference is usually small. But
+there is a CPU/wall time versus size trade off that may warrant investigation.
+
+Output from multi-threaded compression does not require any special handling
+on the decompression side. To the decompressor, data generated with single
+threaded compressor looks the same as data generated by a multi-threaded
+compressor and does not require any special handling or additional resource
+requirements.
+
+Dictionary Creation and Management
+----------------------------------
+
+Compression dictionaries are represented with the ``ZstdCompressionDict`` type.
+
+Instances can be constructed from bytes::
+
+   dict_data = zstd.ZstdCompressionDict(data)
+
+It is possible to construct a dictionary from *any* data. If the data doesn't
+begin with a magic header, it will be treated as a *prefix* dictionary.
+*Prefix* dictionaries allow compression operations to reference raw data
+within the dictionary.
+
+It is possible to force the use of *prefix* dictionaries or to require a
+dictionary header:
+
+   dict_data = zstd.ZstdCompressionDict(data,
+                                        dict_type=zstd.DICT_TYPE_RAWCONTENT)
+
+   dict_data = zstd.ZstdCompressionDict(data,
+                                        dict_type=zstd.DICT_TYPE_FULLDICT)
+
+You can see how many bytes are in the dictionary by calling ``len()``::
+
+   dict_data = zstd.train_dictionary(size, samples)
+   dict_size = len(dict_data)  # will not be larger than ``size``
+
+Once you have a dictionary, you can pass it to the objects performing
+compression and decompression::
+
+   dict_data = zstd.train_dictionary(131072, samples)
+
+   cctx = zstd.ZstdCompressor(dict_data=dict_data)
+   for source_data in input_data:
+       compressed = cctx.compress(source_data)
+       # Do something with compressed data.
+
+   dctx = zstd.ZstdDecompressor(dict_data=dict_data)
+   for compressed_data in input_data:
+       buffer = io.BytesIO()
+       with dctx.stream_writer(buffer) as decompressor:
+           decompressor.write(compressed_data)
+       # Do something with raw data in ``buffer``.
+
+Dictionaries have unique integer IDs. You can retrieve this ID via::
+
+   dict_id = zstd.dictionary_id(dict_data)
+
+You can obtain the raw data in the dict (useful for persisting and constructing
+a ``ZstdCompressionDict`` later) via ``as_bytes()``::
+
+   dict_data = zstd.train_dictionary(size, samples)
+   raw_data = dict_data.as_bytes()
+
+By default, when a ``ZstdCompressionDict`` is *attached* to a
+``ZstdCompressor``, each ``ZstdCompressor`` performs work to prepare the
+dictionary for use. This is fine if only 1 compression operation is being
+performed or if the ``ZstdCompressor`` is being reused for multiple operations.
+But if multiple ``ZstdCompressor`` instances are being used with the dictionary,
+this can add overhead.
+
+It is possible to *precompute* the dictionary so it can readily be consumed
+by multiple ``ZstdCompressor`` instances::
+
+    d = zstd.ZstdCompressionDict(data)
+
+    # Precompute for compression level 3.
+    d.precompute_compress(level=3)
+
+    # Precompute with specific compression parameters.
+    params = zstd.ZstdCompressionParameters(...)
+    d.precompute_compress(compression_params=params)
+
+.. note::
+
+   When a dictionary is precomputed, the compression parameters used to
+   precompute the dictionary overwrite some of the compression parameters
+   specified to ``ZstdCompressor.__init__``.
+
+Training Dictionaries
+^^^^^^^^^^^^^^^^^^^^^
+
+Unless using *prefix* dictionaries, dictionary data is produced by *training*
+on existing data::
+
+   dict_data = zstd.train_dictionary(size, samples)
+
+This takes a target dictionary size and list of bytes instances and creates and
+returns a ``ZstdCompressionDict``.
+
+The dictionary training mechanism is known as *cover*. More details about it are
+available in the paper *Effective Construction of Relative Lempel-Ziv
+Dictionaries* (authors: Liao, Petri, Moffat, Wirth).
+
+The cover algorithm takes parameters ``k` and ``d``. These are the
+*segment size* and *dmer size*, respectively. The returned dictionary
+instance created by this function has ``k`` and ``d`` attributes
+containing the values for these parameters. If a ``ZstdCompressionDict``
+is constructed from raw bytes data (a content-only dictionary), the
+``k`` and ``d`` attributes will be ``0``.
+
+The segment and dmer size parameters to the cover algorithm can either be
+specified manually or ``train_dictionary()`` can try multiple values
+and pick the best one, where *best* means the smallest compressed data size.
+This later mode is called *optimization* mode.
+
+If none of ``k``, ``d``, ``steps``, ``threads``, ``level``, ``notifications``,
+or ``dict_id`` (basically anything from the underlying ``ZDICT_cover_params_t``
+struct) are defined, *optimization* mode is used with default parameter
+values.
+
+If ``steps`` or ``threads`` are defined, then *optimization* mode is engaged
+with explicit control over those parameters. Specifying ``threads=0`` or
+``threads=1`` can be used to engage *optimization* mode if other parameters
+are not defined.
+
+Otherwise, non-*optimization* mode is used with the parameters specified.
+
+This function takes the following arguments:
+
+dict_size
+   Target size in bytes of the dictionary to generate.
+samples
+   A list of bytes holding samples the dictionary will be trained from.
+k
+   Parameter to cover algorithm defining the segment size. A reasonable range
+   is [16, 2048+].
+d
+   Parameter to cover algorithm defining the dmer size. A reasonable range is
+   [6, 16]. ``d`` must be less than or equal to ``k``.
+dict_id
+   Integer dictionary ID for the produced dictionary. Default is 0, which uses
+   a random value.
+steps
+   Number of steps through ``k`` values to perform when trying parameter
+   variations.
+threads
+   Number of threads to use when trying parameter variations. Default is 0,
+   which means to use a single thread. A negative value can be specified to
+   use as many threads as there are detected logical CPUs.
+level
+   Integer target compression level when trying parameter variations.
+notifications
+   Controls writing of informational messages to ``stderr``. ``0`` (the
+   default) means to write nothing. ``1`` writes errors. ``2`` writes
+   progression info. ``3`` writes more details. And ``4`` writes all info.
+
+Explicit Compression Parameters
+-------------------------------
+
+Zstandard offers a high-level *compression level* that maps to lower-level
+compression parameters. For many consumers, this numeric level is the only
+compression setting you'll need to touch.
+
+But for advanced use cases, it might be desirable to tweak these lower-level
+settings.
+
+The ``ZstdCompressionParameters`` type represents these low-level compression
+settings.
+
+Instances of this type can be constructed from a myriad of keyword arguments
+(defined below) for complete low-level control over each adjustable
+compression setting.
+
+From a higher level, one can construct a ``ZstdCompressionParameters`` instance
+given a desired compression level and target input and dictionary size
+using ``ZstdCompressionParameters.from_level()``. e.g.::
+
+    # Derive compression settings for compression level 7.
+    params = zstd.ZstdCompressionParameters.from_level(7)
+
+    # With an input size of 1MB
+    params = zstd.ZstdCompressionParameters.from_level(7, source_size=1048576)
+
+Using ``from_level()``, it is also possible to override individual compression
+parameters or to define additional settings that aren't automatically derived.
+e.g.::
+
+    params = zstd.ZstdCompressionParameters.from_level(4, window_log=10)
+    params = zstd.ZstdCompressionParameters.from_level(5, threads=4)
+
+Or you can define low-level compression settings directly::
+
+    params = zstd.ZstdCompressionParameters(window_log=12, enable_ldm=True)
+
+Once a ``ZstdCompressionParameters`` instance is obtained, it can be used to
+configure a compressor::
+
+    cctx = zstd.ZstdCompressor(compression_params=params)
+
+The named arguments and attributes of ``ZstdCompressionParameters`` are as
+follows:
+
+* format
+* compression_level
+* window_log
+* hash_log
+* chain_log
+* search_log
+* min_match
+* target_length
+* strategy
+* compression_strategy (deprecated: same as ``strategy``)
+* write_content_size
+* write_checksum
+* write_dict_id
+* job_size
+* overlap_log
+* overlap_size_log (deprecated: same as ``overlap_log``)
+* force_max_window
+* enable_ldm
+* ldm_hash_log
+* ldm_min_match
+* ldm_bucket_size_log
+* ldm_hash_rate_log
+* ldm_hash_every_log (deprecated: same as ``ldm_hash_rate_log``)
+* threads
+
+Some of these are very low-level settings. It may help to consult the official
+zstandard documentation for their behavior. Look for the ``ZSTD_p_*`` constants
+in ``zstd.h`` (https://github.com/facebook/zstd/blob/dev/lib/zstd.h).
+
+Frame Inspection
+----------------
+
+Data emitted from zstd compression is encapsulated in a *frame*. This frame
+begins with a 4 byte *magic number* header followed by 2 to 14 bytes describing
+the frame in more detail. For more info, see
+https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md.
+
+``zstd.get_frame_parameters(data)`` parses a zstd *frame* header from a bytes
+instance and return a ``FrameParameters`` object describing the frame.
+
+Depending on which fields are present in the frame and their values, the
+length of the frame parameters varies. If insufficient bytes are passed
+in to fully parse the frame parameters, ``ZstdError`` is raised. To ensure
+frame parameters can be parsed, pass in at least 18 bytes.
+
+``FrameParameters`` instances have the following attributes:
+
+content_size
+   Integer size of original, uncompressed content. This will be ``0`` if the
+   original content size isn't written to the frame (controlled with the
+   ``write_content_size`` argument to ``ZstdCompressor``) or if the input
+   content size was ``0``.
+
+window_size
+   Integer size of maximum back-reference distance in compressed data.
+
+dict_id
+   Integer of dictionary ID used for compression. ``0`` if no dictionary
+   ID was used or if the dictionary ID was ``0``.
+
+has_checksum
+   Bool indicating whether a 4 byte content checksum is stored at the end
+   of the frame.
+
+``zstd.frame_header_size(data)`` returns the size of the zstandard frame
+header.
+
+``zstd.frame_content_size(data)`` returns the content size as parsed from
+the frame header. ``-1`` means the content size is unknown. ``0`` means
+an empty frame. The content size is usually correct. However, it may not
+be accurate.
+
+Misc Functionality
+------------------
+
+estimate_decompression_context_size()
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Estimate the memory size requirements for a decompressor instance.
+
+Constants
+---------
+
+The following module constants/attributes are exposed:
+
+ZSTD_VERSION
+    This module attribute exposes a 3-tuple of the Zstandard version. e.g.
+    ``(1, 0, 0)``
+MAX_COMPRESSION_LEVEL
+    Integer max compression level accepted by compression functions
+COMPRESSION_RECOMMENDED_INPUT_SIZE
+    Recommended chunk size to feed to compressor functions
+COMPRESSION_RECOMMENDED_OUTPUT_SIZE
+    Recommended chunk size for compression output
+DECOMPRESSION_RECOMMENDED_INPUT_SIZE
+    Recommended chunk size to feed into decompresor functions
+DECOMPRESSION_RECOMMENDED_OUTPUT_SIZE
+    Recommended chunk size for decompression output
+
+FRAME_HEADER
+    bytes containing header of the Zstandard frame
+MAGIC_NUMBER
+    Frame header as an integer
+
+FLUSH_BLOCK
+    Flushing behavior that denotes to flush a zstd block. A decompressor will
+    be able to decode all data fed into the compressor so far.
+FLUSH_FRAME
+    Flushing behavior that denotes to end a zstd frame. Any new data fed
+    to the compressor will start a new frame.
+
+CONTENTSIZE_UNKNOWN
+    Value for content size when the content size is unknown.
+CONTENTSIZE_ERROR
+    Value for content size when content size couldn't be determined.
+
+WINDOWLOG_MIN
+    Minimum value for compression parameter
+WINDOWLOG_MAX
+    Maximum value for compression parameter
+CHAINLOG_MIN
+    Minimum value for compression parameter
+CHAINLOG_MAX
+    Maximum value for compression parameter
+HASHLOG_MIN
+    Minimum value for compression parameter
+HASHLOG_MAX
+    Maximum value for compression parameter
+SEARCHLOG_MIN
+    Minimum value for compression parameter
+SEARCHLOG_MAX
+    Maximum value for compression parameter
+MINMATCH_MIN
+    Minimum value for compression parameter
+MINMATCH_MAX
+    Maximum value for compression parameter
+SEARCHLENGTH_MIN
+    Minimum value for compression parameter
+
+    Deprecated: use ``MINMATCH_MIN``
+SEARCHLENGTH_MAX
+    Maximum value for compression parameter
+
+    Deprecated: use ``MINMATCH_MAX``
+TARGETLENGTH_MIN
+    Minimum value for compression parameter
+STRATEGY_FAST
+    Compression strategy
+STRATEGY_DFAST
+    Compression strategy
+STRATEGY_GREEDY
+    Compression strategy
+STRATEGY_LAZY
+    Compression strategy
+STRATEGY_LAZY2
+    Compression strategy
+STRATEGY_BTLAZY2
+    Compression strategy
+STRATEGY_BTOPT
+    Compression strategy
+STRATEGY_BTULTRA
+    Compression strategy
+STRATEGY_BTULTRA2
+    Compression strategy
+
+FORMAT_ZSTD1
+    Zstandard frame format
+FORMAT_ZSTD1_MAGICLESS
+    Zstandard frame format without magic header
+
+Performance Considerations
+--------------------------
+
+The ``ZstdCompressor`` and ``ZstdDecompressor`` types maintain state to a
+persistent compression or decompression *context*. Reusing a ``ZstdCompressor``
+or ``ZstdDecompressor`` instance for multiple operations is faster than
+instantiating a new ``ZstdCompressor`` or ``ZstdDecompressor`` for each
+operation. The differences are magnified as the size of data decreases. For
+example, the difference between *context* reuse and non-reuse for 100,000
+100 byte inputs will be significant (possiby over 10x faster to reuse contexts)
+whereas 10 100,000,000 byte inputs will be more similar in speed (because the
+time spent doing compression dwarfs time spent creating new *contexts*).
+
+Buffer Types
+------------
+
+The API exposes a handful of custom types for interfacing with memory buffers.
+The primary goal of these types is to facilitate efficient multi-object
+operations.
+
+The essential idea is to have a single memory allocation provide backing
+storage for multiple logical objects. This has 2 main advantages: fewer
+allocations and optimal memory access patterns. This avoids having to allocate
+a Python object for each logical object and furthermore ensures that access of
+data for objects can be sequential (read: fast) in memory.
+
+BufferWithSegments
+^^^^^^^^^^^^^^^^^^
+
+The ``BufferWithSegments`` type represents a memory buffer containing N
+discrete items of known lengths (segments). It is essentially a fixed size
+memory address and an array of 2-tuples of ``(offset, length)`` 64-bit
+unsigned native endian integers defining the byte offset and length of each
+segment within the buffer.
+
+Instances behave like containers.
+
+``len()`` returns the number of segments within the instance.
+
+``o[index]`` or ``__getitem__`` obtains a ``BufferSegment`` representing an
+individual segment within the backing buffer. That returned object references
+(not copies) memory. This means that iterating all objects doesn't copy
+data within the buffer.
+
+The ``.size`` attribute contains the total size in bytes of the backing
+buffer.
+
+Instances conform to the buffer protocol. So a reference to the backing bytes
+can be obtained via ``memoryview(o)``. A *copy* of the backing bytes can also
+be obtained via ``.tobytes()``.
+
+The ``.segments`` attribute exposes the array of ``(offset, length)`` for
+segments within the buffer. It is a ``BufferSegments`` type.
+
+BufferSegment
+^^^^^^^^^^^^^
+
+The ``BufferSegment`` type represents a segment within a ``BufferWithSegments``.
+It is essentially a reference to N bytes within a ``BufferWithSegments``.
+
+``len()`` returns the length of the segment in bytes.
+
+``.offset`` contains the byte offset of this segment within its parent
+``BufferWithSegments`` instance.
+
+The object conforms to the buffer protocol. ``.tobytes()`` can be called to
+obtain a ``bytes`` instance with a copy of the backing bytes.
+
+BufferSegments
+^^^^^^^^^^^^^^
+
+This type represents an array of ``(offset, length)`` integers defining segments
+within a ``BufferWithSegments``.
+
+The array members are 64-bit unsigned integers using host/native bit order.
+
+Instances conform to the buffer protocol.
+
+BufferWithSegmentsCollection
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The ``BufferWithSegmentsCollection`` type represents a virtual spanning view
+of multiple ``BufferWithSegments`` instances.
+
+Instances are constructed from 1 or more ``BufferWithSegments`` instances. The
+resulting object behaves like an ordered sequence whose members are the
+segments within each ``BufferWithSegments``.
+
+``len()`` returns the number of segments within all ``BufferWithSegments``
+instances.
+
+``o[index]`` and ``__getitem__(index)`` return the ``BufferSegment`` at
+that offset as if all ``BufferWithSegments`` instances were a single
+entity.
+
+If the object is composed of 2 ``BufferWithSegments`` instances with the
+first having 2 segments and the second have 3 segments, then ``b[0]``
+and ``b[1]`` access segments in the first object and ``b[2]``, ``b[3]``,
+and ``b[4]`` access segments from the second.
+
+Choosing an API
+===============
+
+There are multiple APIs for performing compression and decompression. This is
+because different applications have different needs and the library wants to
+facilitate optimal use in as many use cases as possible.
+
+From a high-level, APIs are divided into *one-shot* and *streaming*: either you
+are operating on all data at once or you operate on it piecemeal.
+
+The *one-shot* APIs are useful for small data, where the input or output
+size is known. (The size can come from a buffer length, file size, or
+stored in the zstd frame header.) A limitation of the *one-shot* APIs is that
+input and output must fit in memory simultaneously. For say a 4 GB input,
+this is often not feasible.
+
+The *one-shot* APIs also perform all work as a single operation. So, if you
+feed it large input, it could take a long time for the function to return.
+
+The streaming APIs do not have the limitations of the simple API. But the
+price you pay for this flexibility is that they are more complex than a
+single function call.
+
+The streaming APIs put the caller in control of compression and decompression
+behavior by allowing them to directly control either the input or output side
+of the operation.
+
+With the *streaming input*, *compressor*, and *decompressor* APIs, the caller
+has full control over the input to the compression or decompression stream.
+They can directly choose when new data is operated on.
+
+With the *streaming ouput* APIs, the caller has full control over the output
+of the compression or decompression stream. It can choose when to receive
+new data.
+
+When using the *streaming* APIs that operate on file-like or stream objects,
+it is important to consider what happens in that object when I/O is requested.
+There is potential for long pauses as data is read or written from the
+underlying stream (say from interacting with a filesystem or network). This
+could add considerable overhead.
+
+Thread Safety
+=============
+
+``ZstdCompressor`` and ``ZstdDecompressor`` instances have no guarantees
+about thread safety. Do not operate on the same ``ZstdCompressor`` and
+``ZstdDecompressor`` instance simultaneously from different threads. It is
+fine to have different threads call into a single instance, just not at the
+same time.
+
+Some operations require multiple function calls to complete. e.g. streaming
+operations. A single ``ZstdCompressor`` or ``ZstdDecompressor`` cannot be used
+for simultaneously active operations. e.g. you must not start a streaming
+operation when another streaming operation is already active.
+
+The C extension releases the GIL during non-trivial calls into the zstd C
+API. Non-trivial calls are notably compression and decompression. Trivial
+calls are things like parsing frame parameters. Where the GIL is released
+is considered an implementation detail and can change in any release.
+
+APIs that accept bytes-like objects don't enforce that the underlying object
+is read-only. However, it is assumed that the passed object is read-only for
+the duration of the function call. It is possible to pass a mutable object
+(like a ``bytearray``) to e.g. ``ZstdCompressor.compress()``, have the GIL
+released, and mutate the object from another thread. Such a race condition
+is a bug in the consumer of python-zstandard. Most Python data types are
+immutable, so unless you are doing something fancy, you don't need to
+worry about this.
+
+Note on Zstandard's *Experimental* API
+======================================
+
+Many of the Zstandard APIs used by this module are marked as *experimental*
+within the Zstandard project.
+
+It is unclear how Zstandard's C API will evolve over time, especially with
+regards to this *experimental* functionality. We will try to maintain
+backwards compatibility at the Python API level. However, we cannot
+guarantee this for things not under our control.
+
+Since a copy of the Zstandard source code is distributed with this
+module and since we compile against it, the behavior of a specific
+version of this module should be constant for all of time. So if you
+pin the version of this module used in your projects (which is a Python
+best practice), you should be shielded from unwanted future changes.
+
+Donate
+======
+
+A lot of time has been invested into this project by the author.
+
+If you find this project useful and would like to thank the author for
+their work, consider donating some money. Any amount is appreciated.
+
+.. image:: https://www.paypalobjects.com/en_US/i/btn/btn_donate_LG.gif
+    :target: https://www.paypal.com/cgi-bin/webscr?cmd=_donations&business=gregory%2eszorc%40gmail%2ecom&lc=US&item_name=python%2dzstandard&currency_code=USD&bn=PP%2dDonationsBF%3abtn_donate_LG%2egif%3aNonHosted
+    :alt: Donate via PayPal
+
+.. |ci-status| image:: https://dev.azure.com/gregoryszorc/python-zstandard/_apis/build/status/indygreg.python-zstandard?branchName=master
+    :target: https://dev.azure.com/gregoryszorc/python-zstandard/_apis/build/status/indygreg.python-zstandard?branchName=master
+
+
diff --git a/contrib/python/zstandard/py2/.dist-info/top_level.txt b/contrib/python/zstandard/py2/.dist-info/top_level.txt
new file mode 100644
index 00000000000..8ed261e4995
--- /dev/null
+++ b/contrib/python/zstandard/py2/.dist-info/top_level.txt
@@ -0,0 +1,2 @@
+zstandard
+zstd
diff --git a/contrib/python/zstandard/py2/LICENSE b/contrib/python/zstandard/py2/LICENSE
new file mode 100644
index 00000000000..dcec4760996
--- /dev/null
+++ b/contrib/python/zstandard/py2/LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2016, Gregory Szorc
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+this list of conditions and the following disclaimer in the documentation
+and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its contributors
+may be used to endorse or promote products derived from this software without
+specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/contrib/python/zstandard/py2/README.rst b/contrib/python/zstandard/py2/README.rst
new file mode 100644
index 00000000000..52bcf32503e
--- /dev/null
+++ b/contrib/python/zstandard/py2/README.rst
@@ -0,0 +1,1615 @@
+================
+python-zstandard
+================
+
+This project provides Python bindings for interfacing with the
+`Zstandard <http://www.zstd.net>`_ compression library. A C extension
+and CFFI interface are provided.
+
+The primary goal of the project is to provide a rich interface to the
+underlying C API through a Pythonic interface while not sacrificing
+performance. This means exposing most of the features and flexibility
+of the C API while not sacrificing usability or safety that Python provides.
+
+The canonical home for this project lives in a Mercurial repository run by
+the author. For convenience, that repository is frequently synchronized to
+https://github.com/indygreg/python-zstandard.
+
+|  |ci-status|
+
+Requirements
+============
+
+This extension is designed to run with Python 2.7, 3.5, 3.6, 3.7, and 3.8
+on common platforms (Linux, Windows, and OS X). On PyPy (both PyPy2 and PyPy3) we support version 6.0.0 and above. 
+x86 and x86_64 are well-tested on Windows. Only x86_64 is well-tested on Linux and macOS.
+
+Installing
+==========
+
+This package is uploaded to PyPI at https://pypi.python.org/pypi/zstandard.
+So, to install this package::
+
+   $ pip install zstandard
+
+Binary wheels are made available for some platforms. If you need to
+install from a source distribution, all you should need is a working C
+compiler and the Python development headers/libraries. On many Linux
+distributions, you can install a ``python-dev`` or ``python-devel``
+package to provide these dependencies.
+
+Packages are also uploaded to Anaconda Cloud at
+https://anaconda.org/indygreg/zstandard. See that URL for how to install
+this package with ``conda``.
+
+Legacy Format Support
+=====================
+
+To enable legacy zstd format support which is needed to handle files compressed
+with zstd < 1.0 you need to provide an installation option::
+
+   $ pip install zstandard --install-option="--legacy"
+
+and since pip 7.0 it is possible to have the following line in your
+requirements.txt::
+
+   zstandard --install-option="--legacy"
+
+Performance
+===========
+
+zstandard is a highly tunable compression algorithm. In its default settings
+(compression level 3), it will be faster at compression and decompression and
+will have better compression ratios than zlib on most data sets. When tuned
+for speed, it approaches lz4's speed and ratios. When tuned for compression
+ratio, it approaches lzma ratios and compression speed, but decompression
+speed is much faster. See the official zstandard documentation for more.
+
+zstandard and this library support multi-threaded compression. There is a
+mechanism to compress large inputs using multiple threads.
+
+The performance of this library is usually very similar to what the zstandard
+C API can deliver. Overhead in this library is due to general Python overhead
+and can't easily be avoided by *any* zstandard Python binding. This library
+exposes multiple APIs for performing compression and decompression so callers
+can pick an API suitable for their need. Contrast with the compression
+modules in Python's standard library (like ``zlib``), which only offer limited
+mechanisms for performing operations. The API flexibility means consumers can
+choose to use APIs that facilitate zero copying or minimize Python object
+creation and garbage collection overhead.
+
+This library is capable of single-threaded throughputs well over 1 GB/s. For
+exact numbers, measure yourself. The source code repository has a ``bench.py``
+script that can be used to measure things.
+
+API
+===
+
+To interface with Zstandard, simply import the ``zstandard`` module::
+
+   import zstandard
+
+It is a popular convention to alias the module as a different name for
+brevity::
+
+   import zstandard as zstd
+
+This module attempts to import and use either the C extension or CFFI
+implementation. On Python platforms known to support C extensions (like
+CPython), it raises an ImportError if the C extension cannot be imported.
+On Python platforms known to not support C extensions (like PyPy), it only
+attempts to import the CFFI implementation and raises ImportError if that
+can't be done. On other platforms, it first tries to import the C extension
+then falls back to CFFI if that fails and raises ImportError if CFFI fails.
+
+To change the module import behavior, a ``PYTHON_ZSTANDARD_IMPORT_POLICY``
+environment variable can be set. The following values are accepted:
+
+default
+   The behavior described above.
+cffi_fallback
+   Always try to import the C extension then fall back to CFFI if that
+   fails.
+cext
+   Only attempt to import the C extension.
+cffi
+   Only attempt to import the CFFI implementation.
+
+In addition, the ``zstandard`` module exports a ``backend`` attribute
+containing the string name of the backend being used. It will be one
+of ``cext`` or ``cffi`` (for *C extension* and *cffi*, respectively).
+
+The types, functions, and attributes exposed by the ``zstandard`` module
+are documented in the sections below.
+
+.. note::
+
+   The documentation in this section makes references to various zstd
+   concepts and functionality. The source repository contains a
+   ``docs/concepts.rst`` file explaining these in more detail.
+
+ZstdCompressor
+--------------
+
+The ``ZstdCompressor`` class provides an interface for performing
+compression operations. Each instance is essentially a wrapper around a
+``ZSTD_CCtx`` from the C API.
+
+Each instance is associated with parameters that control compression
+behavior. These come from the following named arguments (all optional):
+
+level
+   Integer compression level. Valid values are between 1 and 22.
+dict_data
+   Compression dictionary to use.
+
+   Note: When using dictionary data and ``compress()`` is called multiple
+   times, the ``ZstdCompressionParameters`` derived from an integer
+   compression ``level`` and the first compressed data's size will be reused
+   for all subsequent operations. This may not be desirable if source data
+   size varies significantly.
+compression_params
+   A ``ZstdCompressionParameters`` instance defining compression settings.
+write_checksum
+   Whether a 4 byte checksum should be written with the compressed data.
+   Defaults to False. If True, the decompressor can verify that decompressed
+   data matches the original input data.
+write_content_size
+   Whether the size of the uncompressed data will be written into the
+   header of compressed data. Defaults to True. The data will only be
+   written if the compressor knows the size of the input data. This is
+   often not true for streaming compression.
+write_dict_id
+   Whether to write the dictionary ID into the compressed data.
+   Defaults to True. The dictionary ID is only written if a dictionary
+   is being used.
+threads
+   Enables and sets the number of threads to use for multi-threaded compression
+   operations. Defaults to 0, which means to use single-threaded compression.
+   Negative values will resolve to the number of logical CPUs in the system.
+   Read below for more info on multi-threaded compression. This argument only
+   controls thread count for operations that operate on individual pieces of
+   data. APIs that spawn multiple threads for working on multiple pieces of
+   data have their own ``threads`` argument.
+
+``compression_params`` is mutually exclusive with ``level``, ``write_checksum``,
+``write_content_size``, ``write_dict_id``, and ``threads``.
+
+Unless specified otherwise, assume that no two methods of ``ZstdCompressor``
+instances can be called from multiple Python threads simultaneously. In other
+words, assume instances are not thread safe unless stated otherwise.
+
+Utility Methods
+^^^^^^^^^^^^^^^
+
+``frame_progression()`` returns a 3-tuple containing the number of bytes
+ingested, consumed, and produced by the current compression operation.
+
+``memory_size()`` obtains the memory utilization of the underlying zstd
+compression context, in bytes.::
+
+    cctx = zstd.ZstdCompressor()
+    memory = cctx.memory_size()
+
+Simple API
+^^^^^^^^^^
+
+``compress(data)`` compresses and returns data as a one-shot operation.::
+
+   cctx = zstd.ZstdCompressor()
+   compressed = cctx.compress(b'data to compress')
+
+The ``data`` argument can be any object that implements the *buffer protocol*.
+
+Stream Reader API
+^^^^^^^^^^^^^^^^^
+
+``stream_reader(source)`` can be used to obtain an object conforming to the
+``io.RawIOBase`` interface for reading compressed output as a stream::
+
+   with open(path, 'rb') as fh:
+       cctx = zstd.ZstdCompressor()
+       reader = cctx.stream_reader(fh)
+       while True:
+           chunk = reader.read(16384)
+           if not chunk:
+               break
+
+           # Do something with compressed chunk.
+
+Instances can also be used as context managers::
+
+   with open(path, 'rb') as fh:
+       with cctx.stream_reader(fh) as reader:
+           while True:
+               chunk = reader.read(16384)
+               if not chunk:
+                   break
+
+               # Do something with compressed chunk.
+
+When the context manager exits or ``close()`` is called, the stream is closed,
+underlying resources are released, and future operations against the compression
+stream will fail.
+
+The ``source`` argument to ``stream_reader()`` can be any object with a
+``read(size)`` method or any object implementing the *buffer protocol*.
+
+``stream_reader()`` accepts a ``size`` argument specifying how large the input
+stream is. This is used to adjust compression parameters so they are
+tailored to the source size.::
+
+   with open(path, 'rb') as fh:
+       cctx = zstd.ZstdCompressor()
+       with cctx.stream_reader(fh, size=os.stat(path).st_size) as reader:
+           ...
+
+If the ``source`` is a stream, you can specify how large ``read()`` requests
+to that stream should be via the ``read_size`` argument. It defaults to
+``zstandard.COMPRESSION_RECOMMENDED_INPUT_SIZE``.::
+
+   with open(path, 'rb') as fh:
+       cctx = zstd.ZstdCompressor()
+       # Will perform fh.read(8192) when obtaining data to feed into the
+       # compressor.
+       with cctx.stream_reader(fh, read_size=8192) as reader:
+           ...
+
+The stream returned by ``stream_reader()`` is neither writable nor seekable
+(even if the underlying source is seekable). ``readline()`` and
+``readlines()`` are not implemented because they don't make sense for
+compressed data. ``tell()`` returns the number of compressed bytes
+emitted so far.
+
+Streaming Input API
+^^^^^^^^^^^^^^^^^^^
+
+``stream_writer(fh)`` allows you to *stream* data into a compressor.
+
+Returned instances implement the ``io.RawIOBase`` interface. Only methods
+that involve writing will do useful things.
+
+The argument to ``stream_writer()`` must have a ``write(data)`` method. As
+compressed data is available, ``write()`` will be called with the compressed
+data as its argument. Many common Python types implement ``write()``, including
+open file handles and ``io.BytesIO``.
+
+The ``write(data)`` method is used to feed data into the compressor.
+
+The ``flush([flush_mode=FLUSH_BLOCK])`` method can be called to evict whatever
+data remains within the compressor's internal state into the output object. This
+may result in 0 or more ``write()`` calls to the output object. This method
+accepts an optional ``flush_mode`` argument to control the flushing behavior.
+Its value can be any of the ``FLUSH_*`` constants.
+
+Both ``write()`` and ``flush()`` return the number of bytes written to the
+object's ``write()``. In many cases, small inputs do not accumulate enough
+data to cause a write and ``write()`` will return ``0``.
+
+Calling ``close()`` will mark the stream as closed and subsequent I/O
+operations will raise ``ValueError`` (per the documented behavior of
+``io.RawIOBase``). ``close()`` will also call ``close()`` on the underlying
+stream if such a method exists.
+
+Typically usage is as follows::
+
+   cctx = zstd.ZstdCompressor(level=10)
+   compressor = cctx.stream_writer(fh)
+
+   compressor.write(b'chunk 0\n')
+   compressor.write(b'chunk 1\n')
+   compressor.flush()
+   # Receiver will be able to decode ``chunk 0\nchunk 1\n`` at this point.
+   # Receiver is also expecting more data in the zstd *frame*.
+
+   compressor.write(b'chunk 2\n')
+   compressor.flush(zstd.FLUSH_FRAME)
+   # Receiver will be able to decode ``chunk 0\nchunk 1\nchunk 2``.
+   # Receiver is expecting no more data, as the zstd frame is closed.
+   # Any future calls to ``write()`` at this point will construct a new
+   # zstd frame.
+
+Instances can be used as context managers. Exiting the context manager is
+the equivalent of calling ``close()``, which is equivalent to calling
+``flush(zstd.FLUSH_FRAME)``::
+
+   cctx = zstd.ZstdCompressor(level=10)
+   with cctx.stream_writer(fh) as compressor:
+       compressor.write(b'chunk 0')
+       compressor.write(b'chunk 1')
+       ...
+
+.. important::
+
+   If ``flush(FLUSH_FRAME)`` is not called, emitted data doesn't constitute
+   a full zstd *frame* and consumers of this data may complain about malformed
+   input. It is recommended to use instances as a context manager to ensure
+   *frames* are properly finished.
+
+If the size of the data being fed to this streaming compressor is known,
+you can declare it before compression begins::
+
+   cctx = zstd.ZstdCompressor()
+   with cctx.stream_writer(fh, size=data_len) as compressor:
+       compressor.write(chunk0)
+       compressor.write(chunk1)
+       ...
+
+Declaring the size of the source data allows compression parameters to
+be tuned. And if ``write_content_size`` is used, it also results in the
+content size being written into the frame header of the output data.
+
+The size of chunks being ``write()`` to the destination can be specified::
+
+    cctx = zstd.ZstdCompressor()
+    with cctx.stream_writer(fh, write_size=32768) as compressor:
+        ...
+
+To see how much memory is being used by the streaming compressor::
+
+    cctx = zstd.ZstdCompressor()
+    with cctx.stream_writer(fh) as compressor:
+        ...
+        byte_size = compressor.memory_size()
+
+Thte total number of bytes written so far are exposed via ``tell()``::
+
+    cctx = zstd.ZstdCompressor()
+    with cctx.stream_writer(fh) as compressor:
+        ...
+        total_written = compressor.tell()
+
+``stream_writer()`` accepts a ``write_return_read`` boolean argument to control
+the return value of ``write()``. When ``False`` (the default), ``write()`` returns
+the number of bytes that were ``write()``en to the underlying object. When
+``True``, ``write()`` returns the number of bytes read from the input that
+were subsequently written to the compressor. ``True`` is the *proper* behavior
+for ``write()`` as specified by the ``io.RawIOBase`` interface and will become
+the default value in a future release.
+
+Streaming Output API
+^^^^^^^^^^^^^^^^^^^^
+
+``read_to_iter(reader)`` provides a mechanism to stream data out of a
+compressor as an iterator of data chunks.::
+
+   cctx = zstd.ZstdCompressor()
+   for chunk in cctx.read_to_iter(fh):
+        # Do something with emitted data.
+
+``read_to_iter()`` accepts an object that has a ``read(size)`` method or
+conforms to the buffer protocol.
+
+Uncompressed data is fetched from the source either by calling ``read(size)``
+or by fetching a slice of data from the object directly (in the case where
+the buffer protocol is being used). The returned iterator consists of chunks
+of compressed data.
+
+If reading from the source via ``read()``, ``read()`` will be called until
+it raises or returns an empty bytes (``b''``). It is perfectly valid for
+the source to deliver fewer bytes than were what requested by ``read(size)``.
+
+Like ``stream_writer()``, ``read_to_iter()`` also accepts a ``size`` argument
+declaring the size of the input stream::
+
+    cctx = zstd.ZstdCompressor()
+    for chunk in cctx.read_to_iter(fh, size=some_int):
+        pass
+
+You can also control the size that data is ``read()`` from the source and
+the ideal size of output chunks::
+
+    cctx = zstd.ZstdCompressor()
+    for chunk in cctx.read_to_iter(fh, read_size=16384, write_size=8192):
+        pass
+
+Unlike ``stream_writer()``, ``read_to_iter()`` does not give direct control
+over the sizes of chunks fed into the compressor. Instead, chunk sizes will
+be whatever the object being read from delivers. These will often be of a
+uniform size.
+
+Stream Copying API
+^^^^^^^^^^^^^^^^^^
+
+``copy_stream(ifh, ofh)`` can be used to copy data between 2 streams while
+compressing it.::
+
+   cctx = zstd.ZstdCompressor()
+   cctx.copy_stream(ifh, ofh)
+
+For example, say you wish to compress a file::
+
+   cctx = zstd.ZstdCompressor()
+   with open(input_path, 'rb') as ifh, open(output_path, 'wb') as ofh:
+       cctx.copy_stream(ifh, ofh)
+
+It is also possible to declare the size of the source stream::
+
+   cctx = zstd.ZstdCompressor()
+   cctx.copy_stream(ifh, ofh, size=len_of_input)
+
+You can also specify how large the chunks that are ``read()`` and ``write()``
+from and to the streams::
+
+   cctx = zstd.ZstdCompressor()
+   cctx.copy_stream(ifh, ofh, read_size=32768, write_size=16384)
+
+The stream copier returns a 2-tuple of bytes read and written::
+
+   cctx = zstd.ZstdCompressor()
+   read_count, write_count = cctx.copy_stream(ifh, ofh)
+
+Compressor API
+^^^^^^^^^^^^^^
+
+``compressobj()`` returns an object that exposes ``compress(data)`` and
+``flush()`` methods. Each returns compressed data or an empty bytes.
+
+The purpose of ``compressobj()`` is to provide an API-compatible interface
+with ``zlib.compressobj``, ``bz2.BZ2Compressor``, etc. This allows callers to
+swap in different compressor objects while using the same API.
+
+``flush()`` accepts an optional argument indicating how to end the stream.
+``zstd.COMPRESSOBJ_FLUSH_FINISH`` (the default) ends the compression stream.
+Once this type of flush is performed, ``compress()`` and ``flush()`` can
+no longer be called. This type of flush **must** be called to end the
+compression context. If not called, returned data may be incomplete.
+
+A ``zstd.COMPRESSOBJ_FLUSH_BLOCK`` argument to ``flush()`` will flush a
+zstd block. Flushes of this type can be performed multiple times. The next
+call to ``compress()`` will begin a new zstd block.
+
+Here is how this API should be used::
+
+   cctx = zstd.ZstdCompressor()
+   cobj = cctx.compressobj()
+   data = cobj.compress(b'raw input 0')
+   data = cobj.compress(b'raw input 1')
+   data = cobj.flush()
+
+Or to flush blocks::
+
+   cctx.zstd.ZstdCompressor()
+   cobj = cctx.compressobj()
+   data = cobj.compress(b'chunk in first block')
+   data = cobj.flush(zstd.COMPRESSOBJ_FLUSH_BLOCK)
+   data = cobj.compress(b'chunk in second block')
+   data = cobj.flush()
+
+For best performance results, keep input chunks under 256KB. This avoids
+extra allocations for a large output object.
+
+It is possible to declare the input size of the data that will be fed into
+the compressor::
+
+   cctx = zstd.ZstdCompressor()
+   cobj = cctx.compressobj(size=6)
+   data = cobj.compress(b'foobar')
+   data = cobj.flush()
+
+Chunker API
+^^^^^^^^^^^
+
+``chunker(size=None, chunk_size=COMPRESSION_RECOMMENDED_OUTPUT_SIZE)`` returns
+an object that can be used to iteratively feed chunks of data into a compressor
+and produce output chunks of a uniform size.
+
+The object returned by ``chunker()`` exposes the following methods:
+
+``compress(data)``
+   Feeds new input data into the compressor.
+
+``flush()``
+   Flushes all data currently in the compressor.
+
+``finish()``
+   Signals the end of input data. No new data can be compressed after this
+   method is called.
+
+``compress()``, ``flush()``, and ``finish()`` all return an iterator of
+``bytes`` instances holding compressed data. The iterator may be empty. Callers
+MUST iterate through all elements of the returned iterator before performing
+another operation on the object.
+
+All chunks emitted by ``compress()`` will have a length of ``chunk_size``.
+
+``flush()`` and ``finish()`` may return a final chunk smaller than
+``chunk_size``.
+
+Here is how the API should be used::
+
+   cctx = zstd.ZstdCompressor()
+   chunker = cctx.chunker(chunk_size=32768)
+
+   with open(path, 'rb') as fh:
+       while True:
+           in_chunk = fh.read(32768)
+           if not in_chunk:
+               break
+
+           for out_chunk in chunker.compress(in_chunk):
+               # Do something with output chunk of size 32768.
+
+       for out_chunk in chunker.finish():
+           # Do something with output chunks that finalize the zstd frame.
+
+The ``chunker()`` API is often a better alternative to ``compressobj()``.
+
+``compressobj()`` will emit output data as it is available. This results in a
+*stream* of output chunks of varying sizes. The consistency of the output chunk
+size with ``chunker()`` is more appropriate for many usages, such as sending
+compressed data to a socket.
+
+``compressobj()`` may also perform extra memory reallocations in order to
+dynamically adjust the sizes of the output chunks. Since ``chunker()`` output
+chunks are all the same size (except for flushed or final chunks), there is
+less memory allocation overhead.
+
+Batch Compression API
+^^^^^^^^^^^^^^^^^^^^^
+
+(Experimental. Not yet supported in CFFI bindings.)
+
+``multi_compress_to_buffer(data, [threads=0])`` performs compression of multiple
+inputs as a single operation.
+
+Data to be compressed can be passed as a ``BufferWithSegmentsCollection``, a
+``BufferWithSegments``, or a list containing byte like objects. Each element of
+the container will be compressed individually using the configured parameters
+on the ``ZstdCompressor`` instance.
+
+The ``threads`` argument controls how many threads to use for compression. The
+default is ``0`` which means to use a single thread. Negative values use the
+number of logical CPUs in the machine.
+
+The function returns a ``BufferWithSegmentsCollection``. This type represents
+N discrete memory allocations, eaching holding 1 or more compressed frames.
+
+Output data is written to shared memory buffers. This means that unlike
+regular Python objects, a reference to *any* object within the collection
+keeps the shared buffer and therefore memory backing it alive. This can have
+undesirable effects on process memory usage.
+
+The API and behavior of this function is experimental and will likely change.
+Known deficiencies include:
+
+* If asked to use multiple threads, it will always spawn that many threads,
+  even if the input is too small to use them. It should automatically lower
+  the thread count when the extra threads would just add overhead.
+* The buffer allocation strategy is fixed. There is room to make it dynamic,
+  perhaps even to allow one output buffer per input, facilitating a variation
+  of the API to return a list without the adverse effects of shared memory
+  buffers.
+
+ZstdDecompressor
+----------------
+
+The ``ZstdDecompressor`` class provides an interface for performing
+decompression. It is effectively a wrapper around the ``ZSTD_DCtx`` type from
+the C API.
+
+Each instance is associated with parameters that control decompression. These
+come from the following named arguments (all optional):
+
+dict_data
+   Compression dictionary to use.
+max_window_size
+   Sets an uppet limit on the window size for decompression operations in
+   kibibytes. This setting can be used to prevent large memory allocations
+   for inputs using large compression windows.
+format
+   Set the format of data for the decoder. By default, this is
+   ``zstd.FORMAT_ZSTD1``. It can be set to ``zstd.FORMAT_ZSTD1_MAGICLESS`` to
+   allow decoding frames without the 4 byte magic header. Not all decompression
+   APIs support this mode.
+
+The interface of this class is very similar to ``ZstdCompressor`` (by design).
+
+Unless specified otherwise, assume that no two methods of ``ZstdDecompressor``
+instances can be called from multiple Python threads simultaneously. In other
+words, assume instances are not thread safe unless stated otherwise.
+
+Utility Methods
+^^^^^^^^^^^^^^^
+
+``memory_size()`` obtains the size of the underlying zstd decompression context,
+in bytes.::
+
+    dctx = zstd.ZstdDecompressor()
+    size = dctx.memory_size()
+
+Simple API
+^^^^^^^^^^
+
+``decompress(data)`` can be used to decompress an entire compressed zstd
+frame in a single operation.::
+
+    dctx = zstd.ZstdDecompressor()
+    decompressed = dctx.decompress(data)
+
+By default, ``decompress(data)`` will only work on data written with the content
+size encoded in its header (this is the default behavior of
+``ZstdCompressor().compress()`` but may not be true for streaming compression). If
+compressed data without an embedded content size is seen, ``zstd.ZstdError`` will
+be raised.
+
+If the compressed data doesn't have its content size embedded within it,
+decompression can be attempted by specifying the ``max_output_size``
+argument.::
+
+    dctx = zstd.ZstdDecompressor()
+    uncompressed = dctx.decompress(data, max_output_size=1048576)
+
+Ideally, ``max_output_size`` will be identical to the decompressed output
+size.
+
+If ``max_output_size`` is too small to hold the decompressed data,
+``zstd.ZstdError`` will be raised.
+
+If ``max_output_size`` is larger than the decompressed data, the allocated
+output buffer will be resized to only use the space required.
+
+Please note that an allocation of the requested ``max_output_size`` will be
+performed every time the method is called. Setting to a very large value could
+result in a lot of work for the memory allocator and may result in
+``MemoryError`` being raised if the allocation fails.
+
+.. important::
+
+   If the exact size of decompressed data is unknown (not passed in explicitly
+   and not stored in the zstandard frame), for performance reasons it is
+   encouraged to use a streaming API.
+
+Stream Reader API
+^^^^^^^^^^^^^^^^^
+
+``stream_reader(source)`` can be used to obtain an object conforming to the
+``io.RawIOBase`` interface for reading decompressed output as a stream::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       reader = dctx.stream_reader(fh)
+       while True:
+           chunk = reader.read(16384)
+            if not chunk:
+                break
+
+            # Do something with decompressed chunk.
+
+The stream can also be used as a context manager::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       with dctx.stream_reader(fh) as reader:
+           ...
+
+When used as a context manager, the stream is closed and the underlying
+resources are released when the context manager exits. Future operations against
+the stream will fail.
+
+The ``source`` argument to ``stream_reader()`` can be any object with a
+``read(size)`` method or any object implementing the *buffer protocol*.
+
+If the ``source`` is a stream, you can specify how large ``read()`` requests
+to that stream should be via the ``read_size`` argument. It defaults to
+``zstandard.DECOMPRESSION_RECOMMENDED_INPUT_SIZE``.::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       # Will perform fh.read(8192) when obtaining data for the decompressor.
+       with dctx.stream_reader(fh, read_size=8192) as reader:
+           ...
+
+The stream returned by ``stream_reader()`` is not writable.
+
+The stream returned by ``stream_reader()`` is *partially* seekable.
+Absolute and relative positions (``SEEK_SET`` and ``SEEK_CUR``) forward
+of the current position are allowed. Offsets behind the current read
+position and offsets relative to the end of stream are not allowed and
+will raise ``ValueError`` if attempted.
+
+``tell()`` returns the number of decompressed bytes read so far.
+
+Not all I/O methods are implemented. Notably missing is support for
+``readline()``, ``readlines()``, and linewise iteration support. This is
+because streams operate on binary data - not text data. If you want to
+convert decompressed output to text, you can chain an ``io.TextIOWrapper``
+to the stream::
+
+   with open(path, 'rb') as fh:
+       dctx = zstd.ZstdDecompressor()
+       stream_reader = dctx.stream_reader(fh)
+       text_stream = io.TextIOWrapper(stream_reader, encoding='utf-8')
+
+       for line in text_stream:
+           ...
+
+The ``read_across_frames`` argument to ``stream_reader()`` controls the
+behavior of read operations when the end of a zstd *frame* is encountered.
+When ``False`` (the default), a read will complete when the end of a
+zstd *frame* is encountered. When ``True``, a read can potentially
+return data spanning multiple zstd *frames*.
+
+Streaming Input API
+^^^^^^^^^^^^^^^^^^^
+
+``stream_writer(fh)`` allows you to *stream* data into a decompressor.
+
+Returned instances implement the ``io.RawIOBase`` interface. Only methods
+that involve writing will do useful things.
+
+The argument to ``stream_writer()`` is typically an object that also implements
+``io.RawIOBase``. But any object with a ``write(data)`` method will work. Many
+common Python types conform to this interface, including open file handles
+and ``io.BytesIO``.
+
+Behavior is similar to ``ZstdCompressor.stream_writer()``: compressed data
+is sent to the decompressor by calling ``write(data)`` and decompressed
+output is written to the underlying stream by calling its ``write(data)``
+method.::
+
+    dctx = zstd.ZstdDecompressor()
+    decompressor = dctx.stream_writer(fh)
+
+    decompressor.write(compressed_data)
+    ...
+
+
+Calls to ``write()`` will return the number of bytes written to the output
+object. Not all inputs will result in bytes being written, so return values
+of ``0`` are possible.
+
+Like the ``stream_writer()`` compressor, instances can be used as context
+managers. However, context managers add no extra special behavior and offer
+little to no benefit to being used.
+
+Calling ``close()`` will mark the stream as closed and subsequent I/O operations
+will raise ``ValueError`` (per the documented behavior of ``io.RawIOBase``).
+``close()`` will also call ``close()`` on the underlying stream if such a
+method exists.
+
+The size of chunks being ``write()`` to the destination can be specified::
+
+    dctx = zstd.ZstdDecompressor()
+    with dctx.stream_writer(fh, write_size=16384) as decompressor:
+        pass
+
+You can see how much memory is being used by the decompressor::
+
+    dctx = zstd.ZstdDecompressor()
+    with dctx.stream_writer(fh) as decompressor:
+        byte_size = decompressor.memory_size()
+
+``stream_writer()`` accepts a ``write_return_read`` boolean argument to control
+the return value of ``write()``. When ``False`` (the default)``, ``write()``
+returns the number of bytes that were ``write()``en to the underlying stream.
+When ``True``, ``write()`` returns the number of bytes read from the input.
+``True`` is the *proper* behavior for ``write()`` as specified by the
+``io.RawIOBase`` interface and will become the default in a future release.
+
+Streaming Output API
+^^^^^^^^^^^^^^^^^^^^
+
+``read_to_iter(fh)`` provides a mechanism to stream decompressed data out of a
+compressed source as an iterator of data chunks.:: 
+
+    dctx = zstd.ZstdDecompressor()
+    for chunk in dctx.read_to_iter(fh):
+        # Do something with original data.
+
+``read_to_iter()`` accepts an object with a ``read(size)`` method that will
+return  compressed bytes or an object conforming to the buffer protocol that
+can expose its data as a contiguous range of bytes.
+
+``read_to_iter()`` returns an iterator whose elements are chunks of the
+decompressed data.
+
+The size of requested ``read()`` from the source can be specified::
+
+    dctx = zstd.ZstdDecompressor()
+    for chunk in dctx.read_to_iter(fh, read_size=16384):
+        pass
+
+It is also possible to skip leading bytes in the input data::
+
+    dctx = zstd.ZstdDecompressor()
+    for chunk in dctx.read_to_iter(fh, skip_bytes=1):
+        pass
+
+.. tip::
+
+   Skipping leading bytes is useful if the source data contains extra
+   *header* data. Traditionally, you would need to create a slice or
+   ``memoryview`` of the data you want to decompress. This would create
+   overhead. It is more efficient to pass the offset into this API.
+
+Similarly to ``ZstdCompressor.read_to_iter()``, the consumer of the iterator
+controls when data is decompressed. If the iterator isn't consumed,
+decompression is put on hold.
+
+When ``read_to_iter()`` is passed an object conforming to the buffer protocol,
+the behavior may seem similar to what occurs when the simple decompression
+API is used. However, this API works when the decompressed size is unknown.
+Furthermore, if feeding large inputs, the decompressor will work in chunks
+instead of performing a single operation.
+
+Stream Copying API
+^^^^^^^^^^^^^^^^^^
+
+``copy_stream(ifh, ofh)`` can be used to copy data across 2 streams while
+performing decompression.::
+
+    dctx = zstd.ZstdDecompressor()
+    dctx.copy_stream(ifh, ofh)
+
+e.g. to decompress a file to another file::
+
+    dctx = zstd.ZstdDecompressor()
+    with open(input_path, 'rb') as ifh, open(output_path, 'wb') as ofh:
+        dctx.copy_stream(ifh, ofh)
+
+The size of chunks being ``read()`` and ``write()`` from and to the streams
+can be specified::
+
+    dctx = zstd.ZstdDecompressor()
+    dctx.copy_stream(ifh, ofh, read_size=8192, write_size=16384)
+
+Decompressor API
+^^^^^^^^^^^^^^^^
+
+``decompressobj()`` returns an object that exposes a ``decompress(data)``
+method. Compressed data chunks are fed into ``decompress(data)`` and
+uncompressed output (or an empty bytes) is returned. Output from subsequent
+calls needs to be concatenated to reassemble the full decompressed byte
+sequence.
+
+The purpose of ``decompressobj()`` is to provide an API-compatible interface
+with ``zlib.decompressobj`` and ``bz2.BZ2Decompressor``. This allows callers
+to swap in different decompressor objects while using the same API.
+
+Each object is single use: once an input frame is decoded, ``decompress()``
+can no longer be called.
+
+Here is how this API should be used::
+
+   dctx = zstd.ZstdDecompressor()
+   dobj = dctx.decompressobj()
+   data = dobj.decompress(compressed_chunk_0)
+   data = dobj.decompress(compressed_chunk_1)
+
+By default, calls to ``decompress()`` write output data in chunks of size
+``DECOMPRESSION_RECOMMENDED_OUTPUT_SIZE``. These chunks are concatenated
+before being returned to the caller. It is possible to define the size of
+these temporary chunks by passing ``write_size`` to ``decompressobj()``::
+
+   dctx = zstd.ZstdDecompressor()
+   dobj = dctx.decompressobj(write_size=1048576)
+
+.. note::
+
+   Because calls to ``decompress()`` may need to perform multiple
+   memory (re)allocations, this streaming decompression API isn't as
+   efficient as other APIs.
+
+For compatibility with the standard library APIs, instances expose a
+``flush([length=None])`` method. This method no-ops and has no meaningful
+side-effects, making it safe to call any time.
+
+Batch Decompression API
+^^^^^^^^^^^^^^^^^^^^^^^
+
+(Experimental. Not yet supported in CFFI bindings.)
+
+``multi_decompress_to_buffer()`` performs decompression of multiple
+frames as a single operation and returns a ``BufferWithSegmentsCollection``
+containing decompressed data for all inputs.
+
+Compressed frames can be passed to the function as a ``BufferWithSegments``,
+a ``BufferWithSegmentsCollection``, or as a list containing objects that
+conform to the buffer protocol. For best performance, pass a
+``BufferWithSegmentsCollection`` or a ``BufferWithSegments``, as
+minimal input validation will be done for that type. If calling from
+Python (as opposed to C), constructing one of these instances may add
+overhead cancelling out the performance overhead of validation for list
+inputs.::
+
+    dctx = zstd.ZstdDecompressor()
+    results = dctx.multi_decompress_to_buffer([b'...', b'...'])
+
+The decompressed size of each frame MUST be discoverable. It can either be
+embedded within the zstd frame (``write_content_size=True`` argument to
+``ZstdCompressor``) or passed in via the ``decompressed_sizes`` argument.
+
+The ``decompressed_sizes`` argument is an object conforming to the buffer
+protocol which holds an array of 64-bit unsigned integers in the machine's
+native format defining the decompressed sizes of each frame. If this argument
+is passed, it avoids having to scan each frame for its decompressed size.
+This frame scanning can add noticeable overhead in some scenarios.::
+
+    frames = [...]
+    sizes = struct.pack('=QQQQ', len0, len1, len2, len3)
+
+    dctx = zstd.ZstdDecompressor()
+    results = dctx.multi_decompress_to_buffer(frames, decompressed_sizes=sizes)
+
+The ``threads`` argument controls the number of threads to use to perform
+decompression operations. The default (``0``) or the value ``1`` means to
+use a single thread. Negative values use the number of logical CPUs in the
+machine.
+
+.. note::
+
+   It is possible to pass a ``mmap.mmap()`` instance into this function by
+   wrapping it with a ``BufferWithSegments`` instance (which will define the
+   offsets of frames within the memory mapped region).
+
+This function is logically equivalent to performing ``dctx.decompress()``
+on each input frame and returning the result.
+
+This function exists to perform decompression on multiple frames as fast
+as possible by having as little overhead as possible. Since decompression is
+performed as a single operation and since the decompressed output is stored in
+a single buffer, extra memory allocations, Python objects, and Python function
+calls are avoided. This is ideal for scenarios where callers know up front that
+they need to access data for multiple frames, such as when  *delta chains* are
+being used.
+
+Currently, the implementation always spawns multiple threads when requested,
+even if the amount of work to do is small. In the future, it will be smarter
+about avoiding threads and their associated overhead when the amount of
+work to do is small.
+
+Prefix Dictionary Chain Decompression
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``decompress_content_dict_chain(frames)`` performs decompression of a list of
+zstd frames produced using chained *prefix* dictionary compression. Such
+a list of frames is produced by compressing discrete inputs where each
+non-initial input is compressed with a *prefix* dictionary consisting of the
+content of the previous input.
+
+For example, say you have the following inputs::
+
+   inputs = [b'input 1', b'input 2', b'input 3']
+
+The zstd frame chain consists of:
+
+1. ``b'input 1'`` compressed in standalone/discrete mode
+2. ``b'input 2'`` compressed using ``b'input 1'`` as a *prefix* dictionary
+3. ``b'input 3'`` compressed using ``b'input 2'`` as a *prefix* dictionary
+
+Each zstd frame **must** have the content size written.
+
+The following Python code can be used to produce a *prefix dictionary chain*::
+
+    def make_chain(inputs):
+        frames = []
+
+        # First frame is compressed in standalone/discrete mode.
+        zctx = zstd.ZstdCompressor()
+        frames.append(zctx.compress(inputs[0]))
+
+        # Subsequent frames use the previous fulltext as a prefix dictionary
+        for i, raw in enumerate(inputs[1:]):
+            dict_data = zstd.ZstdCompressionDict(
+                inputs[i], dict_type=zstd.DICT_TYPE_RAWCONTENT)
+            zctx = zstd.ZstdCompressor(dict_data=dict_data)
+            frames.append(zctx.compress(raw))
+
+        return frames
+
+``decompress_content_dict_chain()`` returns the uncompressed data of the last
+element in the input chain.
+
+
+.. note::
+
+   It is possible to implement *prefix dictionary chain* decompression
+   on top of other APIs. However, this function will likely be faster -
+   especially for long input chains - as it avoids the overhead of instantiating
+   and passing around intermediate objects between C and Python.
+
+Multi-Threaded Compression
+--------------------------
+
+``ZstdCompressor`` accepts a ``threads`` argument that controls the number
+of threads to use for compression. The way this works is that input is split
+into segments and each segment is fed into a worker pool for compression. Once
+a segment is compressed, it is flushed/appended to the output.
+
+.. note::
+
+   These threads are created at the C layer and are not Python threads. So they
+   work outside the GIL. It is therefore possible to CPU saturate multiple cores
+   from Python.
+
+The segment size for multi-threaded compression is chosen from the window size
+of the compressor. This is derived from the ``window_log`` attribute of a
+``ZstdCompressionParameters`` instance. By default, segment sizes are in the 1+MB
+range.
+
+If multi-threaded compression is requested and the input is smaller than the
+configured segment size, only a single compression thread will be used. If the
+input is smaller than the segment size multiplied by the thread pool size or
+if data cannot be delivered to the compressor fast enough, not all requested
+compressor threads may be active simultaneously.
+
+Compared to non-multi-threaded compression, multi-threaded compression has
+higher per-operation overhead. This includes extra memory operations,
+thread creation, lock acquisition, etc.
+
+Due to the nature of multi-threaded compression using *N* compression
+*states*, the output from multi-threaded compression will likely be larger
+than non-multi-threaded compression. The difference is usually small. But
+there is a CPU/wall time versus size trade off that may warrant investigation.
+
+Output from multi-threaded compression does not require any special handling
+on the decompression side. To the decompressor, data generated with single
+threaded compressor looks the same as data generated by a multi-threaded
+compressor and does not require any special handling or additional resource
+requirements.
+
+Dictionary Creation and Management
+----------------------------------
+
+Compression dictionaries are represented with the ``ZstdCompressionDict`` type.
+
+Instances can be constructed from bytes::
+
+   dict_data = zstd.ZstdCompressionDict(data)
+
+It is possible to construct a dictionary from *any* data. If the data doesn't
+begin with a magic header, it will be treated as a *prefix* dictionary.
+*Prefix* dictionaries allow compression operations to reference raw data
+within the dictionary.
+
+It is possible to force the use of *prefix* dictionaries or to require a
+dictionary header:
+
+   dict_data = zstd.ZstdCompressionDict(data,
+                                        dict_type=zstd.DICT_TYPE_RAWCONTENT)
+
+   dict_data = zstd.ZstdCompressionDict(data,
+                                        dict_type=zstd.DICT_TYPE_FULLDICT)
+
+You can see how many bytes are in the dictionary by calling ``len()``::
+
+   dict_data = zstd.train_dictionary(size, samples)
+   dict_size = len(dict_data)  # will not be larger than ``size``
+
+Once you have a dictionary, you can pass it to the objects performing
+compression and decompression::
+
+   dict_data = zstd.train_dictionary(131072, samples)
+
+   cctx = zstd.ZstdCompressor(dict_data=dict_data)
+   for source_data in input_data:
+       compressed = cctx.compress(source_data)
+       # Do something with compressed data.
+
+   dctx = zstd.ZstdDecompressor(dict_data=dict_data)
+   for compressed_data in input_data:
+       buffer = io.BytesIO()
+       with dctx.stream_writer(buffer) as decompressor:
+           decompressor.write(compressed_data)
+       # Do something with raw data in ``buffer``.
+
+Dictionaries have unique integer IDs. You can retrieve this ID via::
+
+   dict_id = zstd.dictionary_id(dict_data)
+
+You can obtain the raw data in the dict (useful for persisting and constructing
+a ``ZstdCompressionDict`` later) via ``as_bytes()``::
+
+   dict_data = zstd.train_dictionary(size, samples)
+   raw_data = dict_data.as_bytes()
+
+By default, when a ``ZstdCompressionDict`` is *attached* to a
+``ZstdCompressor``, each ``ZstdCompressor`` performs work to prepare the
+dictionary for use. This is fine if only 1 compression operation is being
+performed or if the ``ZstdCompressor`` is being reused for multiple operations.
+But if multiple ``ZstdCompressor`` instances are being used with the dictionary,
+this can add overhead.
+
+It is possible to *precompute* the dictionary so it can readily be consumed
+by multiple ``ZstdCompressor`` instances::
+
+    d = zstd.ZstdCompressionDict(data)
+
+    # Precompute for compression level 3.
+    d.precompute_compress(level=3)
+
+    # Precompute with specific compression parameters.
+    params = zstd.ZstdCompressionParameters(...)
+    d.precompute_compress(compression_params=params)
+
+.. note::
+
+   When a dictionary is precomputed, the compression parameters used to
+   precompute the dictionary overwrite some of the compression parameters
+   specified to ``ZstdCompressor.__init__``.
+
+Training Dictionaries
+^^^^^^^^^^^^^^^^^^^^^
+
+Unless using *prefix* dictionaries, dictionary data is produced by *training*
+on existing data::
+
+   dict_data = zstd.train_dictionary(size, samples)
+
+This takes a target dictionary size and list of bytes instances and creates and
+returns a ``ZstdCompressionDict``.
+
+The dictionary training mechanism is known as *cover*. More details about it are
+available in the paper *Effective Construction of Relative Lempel-Ziv
+Dictionaries* (authors: Liao, Petri, Moffat, Wirth).
+
+The cover algorithm takes parameters ``k` and ``d``. These are the
+*segment size* and *dmer size*, respectively. The returned dictionary
+instance created by this function has ``k`` and ``d`` attributes
+containing the values for these parameters. If a ``ZstdCompressionDict``
+is constructed from raw bytes data (a content-only dictionary), the
+``k`` and ``d`` attributes will be ``0``.
+
+The segment and dmer size parameters to the cover algorithm can either be
+specified manually or ``train_dictionary()`` can try multiple values
+and pick the best one, where *best* means the smallest compressed data size.
+This later mode is called *optimization* mode.
+
+If none of ``k``, ``d``, ``steps``, ``threads``, ``level``, ``notifications``,
+or ``dict_id`` (basically anything from the underlying ``ZDICT_cover_params_t``
+struct) are defined, *optimization* mode is used with default parameter
+values.
+
+If ``steps`` or ``threads`` are defined, then *optimization* mode is engaged
+with explicit control over those parameters. Specifying ``threads=0`` or
+``threads=1`` can be used to engage *optimization* mode if other parameters
+are not defined.
+
+Otherwise, non-*optimization* mode is used with the parameters specified.
+
+This function takes the following arguments:
+
+dict_size
+   Target size in bytes of the dictionary to generate.
+samples
+   A list of bytes holding samples the dictionary will be trained from.
+k
+   Parameter to cover algorithm defining the segment size. A reasonable range
+   is [16, 2048+].
+d
+   Parameter to cover algorithm defining the dmer size. A reasonable range is
+   [6, 16]. ``d`` must be less than or equal to ``k``.
+dict_id
+   Integer dictionary ID for the produced dictionary. Default is 0, which uses
+   a random value.
+steps
+   Number of steps through ``k`` values to perform when trying parameter
+   variations.
+threads
+   Number of threads to use when trying parameter variations. Default is 0,
+   which means to use a single thread. A negative value can be specified to
+   use as many threads as there are detected logical CPUs.
+level
+   Integer target compression level when trying parameter variations.
+notifications
+   Controls writing of informational messages to ``stderr``. ``0`` (the
+   default) means to write nothing. ``1`` writes errors. ``2`` writes
+   progression info. ``3`` writes more details. And ``4`` writes all info.
+
+Explicit Compression Parameters
+-------------------------------
+
+Zstandard offers a high-level *compression level* that maps to lower-level
+compression parameters. For many consumers, this numeric level is the only
+compression setting you'll need to touch.
+
+But for advanced use cases, it might be desirable to tweak these lower-level
+settings.
+
+The ``ZstdCompressionParameters`` type represents these low-level compression
+settings.
+
+Instances of this type can be constructed from a myriad of keyword arguments
+(defined below) for complete low-level control over each adjustable
+compression setting.
+
+From a higher level, one can construct a ``ZstdCompressionParameters`` instance
+given a desired compression level and target input and dictionary size
+using ``ZstdCompressionParameters.from_level()``. e.g.::
+
+    # Derive compression settings for compression level 7.
+    params = zstd.ZstdCompressionParameters.from_level(7)
+
+    # With an input size of 1MB
+    params = zstd.ZstdCompressionParameters.from_level(7, source_size=1048576)
+
+Using ``from_level()``, it is also possible to override individual compression
+parameters or to define additional settings that aren't automatically derived.
+e.g.::
+
+    params = zstd.ZstdCompressionParameters.from_level(4, window_log=10)
+    params = zstd.ZstdCompressionParameters.from_level(5, threads=4)
+
+Or you can define low-level compression settings directly::
+
+    params = zstd.ZstdCompressionParameters(window_log=12, enable_ldm=True)
+
+Once a ``ZstdCompressionParameters`` instance is obtained, it can be used to
+configure a compressor::
+
+    cctx = zstd.ZstdCompressor(compression_params=params)
+
+The named arguments and attributes of ``ZstdCompressionParameters`` are as
+follows:
+
+* format
+* compression_level
+* window_log
+* hash_log
+* chain_log
+* search_log
+* min_match
+* target_length
+* strategy
+* compression_strategy (deprecated: same as ``strategy``)
+* write_content_size
+* write_checksum
+* write_dict_id
+* job_size
+* overlap_log
+* overlap_size_log (deprecated: same as ``overlap_log``)
+* force_max_window
+* enable_ldm
+* ldm_hash_log
+* ldm_min_match
+* ldm_bucket_size_log
+* ldm_hash_rate_log
+* ldm_hash_every_log (deprecated: same as ``ldm_hash_rate_log``)
+* threads
+
+Some of these are very low-level settings. It may help to consult the official
+zstandard documentation for their behavior. Look for the ``ZSTD_p_*`` constants
+in ``zstd.h`` (https://github.com/facebook/zstd/blob/dev/lib/zstd.h).
+
+Frame Inspection
+----------------
+
+Data emitted from zstd compression is encapsulated in a *frame*. This frame
+begins with a 4 byte *magic number* header followed by 2 to 14 bytes describing
+the frame in more detail. For more info, see
+https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md.
+
+``zstd.get_frame_parameters(data)`` parses a zstd *frame* header from a bytes
+instance and return a ``FrameParameters`` object describing the frame.
+
+Depending on which fields are present in the frame and their values, the
+length of the frame parameters varies. If insufficient bytes are passed
+in to fully parse the frame parameters, ``ZstdError`` is raised. To ensure
+frame parameters can be parsed, pass in at least 18 bytes.
+
+``FrameParameters`` instances have the following attributes:
+
+content_size
+   Integer size of original, uncompressed content. This will be ``0`` if the
+   original content size isn't written to the frame (controlled with the
+   ``write_content_size`` argument to ``ZstdCompressor``) or if the input
+   content size was ``0``.
+
+window_size
+   Integer size of maximum back-reference distance in compressed data.
+
+dict_id
+   Integer of dictionary ID used for compression. ``0`` if no dictionary
+   ID was used or if the dictionary ID was ``0``.
+
+has_checksum
+   Bool indicating whether a 4 byte content checksum is stored at the end
+   of the frame.
+
+``zstd.frame_header_size(data)`` returns the size of the zstandard frame
+header.
+
+``zstd.frame_content_size(data)`` returns the content size as parsed from
+the frame header. ``-1`` means the content size is unknown. ``0`` means
+an empty frame. The content size is usually correct. However, it may not
+be accurate.
+
+Misc Functionality
+------------------
+
+estimate_decompression_context_size()
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Estimate the memory size requirements for a decompressor instance.
+
+Constants
+---------
+
+The following module constants/attributes are exposed:
+
+ZSTD_VERSION
+    This module attribute exposes a 3-tuple of the Zstandard version. e.g.
+    ``(1, 0, 0)``
+MAX_COMPRESSION_LEVEL
+    Integer max compression level accepted by compression functions
+COMPRESSION_RECOMMENDED_INPUT_SIZE
+    Recommended chunk size to feed to compressor functions
+COMPRESSION_RECOMMENDED_OUTPUT_SIZE
+    Recommended chunk size for compression output
+DECOMPRESSION_RECOMMENDED_INPUT_SIZE
+    Recommended chunk size to feed into decompresor functions
+DECOMPRESSION_RECOMMENDED_OUTPUT_SIZE
+    Recommended chunk size for decompression output
+
+FRAME_HEADER
+    bytes containing header of the Zstandard frame
+MAGIC_NUMBER
+    Frame header as an integer
+
+FLUSH_BLOCK
+    Flushing behavior that denotes to flush a zstd block. A decompressor will
+    be able to decode all data fed into the compressor so far.
+FLUSH_FRAME
+    Flushing behavior that denotes to end a zstd frame. Any new data fed
+    to the compressor will start a new frame.
+
+CONTENTSIZE_UNKNOWN
+    Value for content size when the content size is unknown.
+CONTENTSIZE_ERROR
+    Value for content size when content size couldn't be determined.
+
+WINDOWLOG_MIN
+    Minimum value for compression parameter
+WINDOWLOG_MAX
+    Maximum value for compression parameter
+CHAINLOG_MIN
+    Minimum value for compression parameter
+CHAINLOG_MAX
+    Maximum value for compression parameter
+HASHLOG_MIN
+    Minimum value for compression parameter
+HASHLOG_MAX
+    Maximum value for compression parameter
+SEARCHLOG_MIN
+    Minimum value for compression parameter
+SEARCHLOG_MAX
+    Maximum value for compression parameter
+MINMATCH_MIN
+    Minimum value for compression parameter
+MINMATCH_MAX
+    Maximum value for compression parameter
+SEARCHLENGTH_MIN
+    Minimum value for compression parameter
+
+    Deprecated: use ``MINMATCH_MIN``
+SEARCHLENGTH_MAX
+    Maximum value for compression parameter
+
+    Deprecated: use ``MINMATCH_MAX``
+TARGETLENGTH_MIN
+    Minimum value for compression parameter
+STRATEGY_FAST
+    Compression strategy
+STRATEGY_DFAST
+    Compression strategy
+STRATEGY_GREEDY
+    Compression strategy
+STRATEGY_LAZY
+    Compression strategy
+STRATEGY_LAZY2
+    Compression strategy
+STRATEGY_BTLAZY2
+    Compression strategy
+STRATEGY_BTOPT
+    Compression strategy
+STRATEGY_BTULTRA
+    Compression strategy
+STRATEGY_BTULTRA2
+    Compression strategy
+
+FORMAT_ZSTD1
+    Zstandard frame format
+FORMAT_ZSTD1_MAGICLESS
+    Zstandard frame format without magic header
+
+Performance Considerations
+--------------------------
+
+The ``ZstdCompressor`` and ``ZstdDecompressor`` types maintain state to a
+persistent compression or decompression *context*. Reusing a ``ZstdCompressor``
+or ``ZstdDecompressor`` instance for multiple operations is faster than
+instantiating a new ``ZstdCompressor`` or ``ZstdDecompressor`` for each
+operation. The differences are magnified as the size of data decreases. For
+example, the difference between *context* reuse and non-reuse for 100,000
+100 byte inputs will be significant (possiby over 10x faster to reuse contexts)
+whereas 10 100,000,000 byte inputs will be more similar in speed (because the
+time spent doing compression dwarfs time spent creating new *contexts*).
+
+Buffer Types
+------------
+
+The API exposes a handful of custom types for interfacing with memory buffers.
+The primary goal of these types is to facilitate efficient multi-object
+operations.
+
+The essential idea is to have a single memory allocation provide backing
+storage for multiple logical objects. This has 2 main advantages: fewer
+allocations and optimal memory access patterns. This avoids having to allocate
+a Python object for each logical object and furthermore ensures that access of
+data for objects can be sequential (read: fast) in memory.
+
+BufferWithSegments
+^^^^^^^^^^^^^^^^^^
+
+The ``BufferWithSegments`` type represents a memory buffer containing N
+discrete items of known lengths (segments). It is essentially a fixed size
+memory address and an array of 2-tuples of ``(offset, length)`` 64-bit
+unsigned native endian integers defining the byte offset and length of each
+segment within the buffer.
+
+Instances behave like containers.
+
+``len()`` returns the number of segments within the instance.
+
+``o[index]`` or ``__getitem__`` obtains a ``BufferSegment`` representing an
+individual segment within the backing buffer. That returned object references
+(not copies) memory. This means that iterating all objects doesn't copy
+data within the buffer.
+
+The ``.size`` attribute contains the total size in bytes of the backing
+buffer.
+
+Instances conform to the buffer protocol. So a reference to the backing bytes
+can be obtained via ``memoryview(o)``. A *copy* of the backing bytes can also
+be obtained via ``.tobytes()``.
+
+The ``.segments`` attribute exposes the array of ``(offset, length)`` for
+segments within the buffer. It is a ``BufferSegments`` type.
+
+BufferSegment
+^^^^^^^^^^^^^
+
+The ``BufferSegment`` type represents a segment within a ``BufferWithSegments``.
+It is essentially a reference to N bytes within a ``BufferWithSegments``.
+
+``len()`` returns the length of the segment in bytes.
+
+``.offset`` contains the byte offset of this segment within its parent
+``BufferWithSegments`` instance.
+
+The object conforms to the buffer protocol. ``.tobytes()`` can be called to
+obtain a ``bytes`` instance with a copy of the backing bytes.
+
+BufferSegments
+^^^^^^^^^^^^^^
+
+This type represents an array of ``(offset, length)`` integers defining segments
+within a ``BufferWithSegments``.
+
+The array members are 64-bit unsigned integers using host/native bit order.
+
+Instances conform to the buffer protocol.
+
+BufferWithSegmentsCollection
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The ``BufferWithSegmentsCollection`` type represents a virtual spanning view
+of multiple ``BufferWithSegments`` instances.
+
+Instances are constructed from 1 or more ``BufferWithSegments`` instances. The
+resulting object behaves like an ordered sequence whose members are the
+segments within each ``BufferWithSegments``.
+
+``len()`` returns the number of segments within all ``BufferWithSegments``
+instances.
+
+``o[index]`` and ``__getitem__(index)`` return the ``BufferSegment`` at
+that offset as if all ``BufferWithSegments`` instances were a single
+entity.
+
+If the object is composed of 2 ``BufferWithSegments`` instances with the
+first having 2 segments and the second have 3 segments, then ``b[0]``
+and ``b[1]`` access segments in the first object and ``b[2]``, ``b[3]``,
+and ``b[4]`` access segments from the second.
+
+Choosing an API
+===============
+
+There are multiple APIs for performing compression and decompression. This is
+because different applications have different needs and the library wants to
+facilitate optimal use in as many use cases as possible.
+
+From a high-level, APIs are divided into *one-shot* and *streaming*: either you
+are operating on all data at once or you operate on it piecemeal.
+
+The *one-shot* APIs are useful for small data, where the input or output
+size is known. (The size can come from a buffer length, file size, or
+stored in the zstd frame header.) A limitation of the *one-shot* APIs is that
+input and output must fit in memory simultaneously. For say a 4 GB input,
+this is often not feasible.
+
+The *one-shot* APIs also perform all work as a single operation. So, if you
+feed it large input, it could take a long time for the function to return.
+
+The streaming APIs do not have the limitations of the simple API. But the
+price you pay for this flexibility is that they are more complex than a
+single function call.
+
+The streaming APIs put the caller in control of compression and decompression
+behavior by allowing them to directly control either the input or output side
+of the operation.
+
+With the *streaming input*, *compressor*, and *decompressor* APIs, the caller
+has full control over the input to the compression or decompression stream.
+They can directly choose when new data is operated on.
+
+With the *streaming ouput* APIs, the caller has full control over the output
+of the compression or decompression stream. It can choose when to receive
+new data.
+
+When using the *streaming* APIs that operate on file-like or stream objects,
+it is important to consider what happens in that object when I/O is requested.
+There is potential for long pauses as data is read or written from the
+underlying stream (say from interacting with a filesystem or network). This
+could add considerable overhead.
+
+Thread Safety
+=============
+
+``ZstdCompressor`` and ``ZstdDecompressor`` instances have no guarantees
+about thread safety. Do not operate on the same ``ZstdCompressor`` and
+``ZstdDecompressor`` instance simultaneously from different threads. It is
+fine to have different threads call into a single instance, just not at the
+same time.
+
+Some operations require multiple function calls to complete. e.g. streaming
+operations. A single ``ZstdCompressor`` or ``ZstdDecompressor`` cannot be used
+for simultaneously active operations. e.g. you must not start a streaming
+operation when another streaming operation is already active.
+
+The C extension releases the GIL during non-trivial calls into the zstd C
+API. Non-trivial calls are notably compression and decompression. Trivial
+calls are things like parsing frame parameters. Where the GIL is released
+is considered an implementation detail and can change in any release.
+
+APIs that accept bytes-like objects don't enforce that the underlying object
+is read-only. However, it is assumed that the passed object is read-only for
+the duration of the function call. It is possible to pass a mutable object
+(like a ``bytearray``) to e.g. ``ZstdCompressor.compress()``, have the GIL
+released, and mutate the object from another thread. Such a race condition
+is a bug in the consumer of python-zstandard. Most Python data types are
+immutable, so unless you are doing something fancy, you don't need to
+worry about this.
+
+Note on Zstandard's *Experimental* API
+======================================
+
+Many of the Zstandard APIs used by this module are marked as *experimental*
+within the Zstandard project.
+
+It is unclear how Zstandard's C API will evolve over time, especially with
+regards to this *experimental* functionality. We will try to maintain
+backwards compatibility at the Python API level. However, we cannot
+guarantee this for things not under our control.
+
+Since a copy of the Zstandard source code is distributed with this
+module and since we compile against it, the behavior of a specific
+version of this module should be constant for all of time. So if you
+pin the version of this module used in your projects (which is a Python
+best practice), you should be shielded from unwanted future changes.
+
+Donate
+======
+
+A lot of time has been invested into this project by the author.
+
+If you find this project useful and would like to thank the author for
+their work, consider donating some money. Any amount is appreciated.
+
+.. image:: https://www.paypalobjects.com/en_US/i/btn/btn_donate_LG.gif
+    :target: https://www.paypal.com/cgi-bin/webscr?cmd=_donations&business=gregory%2eszorc%40gmail%2ecom&lc=US&item_name=python%2dzstandard&currency_code=USD&bn=PP%2dDonationsBF%3abtn_donate_LG%2egif%3aNonHosted
+    :alt: Donate via PayPal
+
+.. |ci-status| image:: https://dev.azure.com/gregoryszorc/python-zstandard/_apis/build/status/indygreg.python-zstandard?branchName=master
+    :target: https://dev.azure.com/gregoryszorc/python-zstandard/_apis/build/status/indygreg.python-zstandard?branchName=master
diff --git a/contrib/python/zstandard/py2/c-ext/bufferutil.c b/contrib/python/zstandard/py2/c-ext/bufferutil.c
new file mode 100644
index 00000000000..5094a4bb92f
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/bufferutil.c
@@ -0,0 +1,792 @@
+/**
+* Copyright (c) 2017-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(BufferWithSegments__doc__,
+"BufferWithSegments - A memory buffer holding known sub-segments.\n"
+"\n"
+"This type represents a contiguous chunk of memory containing N discrete\n"
+"items within sub-segments of that memory.\n"
+"\n"
+"Segments within the buffer are stored as an array of\n"
+"``(offset, length)`` pairs, where each element is an unsigned 64-bit\n"
+"integer using the host/native bit order representation.\n"
+"\n"
+"The type exists to facilitate operations against N>1 items without the\n"
+"overhead of Python object creation and management.\n"
+);
+
+static void BufferWithSegments_dealloc(ZstdBufferWithSegments* self) {
+	/* Backing memory is either canonically owned by a Py_buffer or by us. */
+	if (self->parent.buf) {
+		PyBuffer_Release(&self->parent);
+	}
+	else if (self->useFree) {
+		free(self->data);
+	}
+	else {
+		PyMem_Free(self->data);
+	}
+
+	self->data = NULL;
+
+	if (self->useFree) {
+		free(self->segments);
+	}
+	else {
+		PyMem_Free(self->segments);
+	}
+
+	self->segments = NULL;
+
+	PyObject_Del(self);
+}
+
+static int BufferWithSegments_init(ZstdBufferWithSegments* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		"segments",
+		NULL
+	};
+
+	Py_buffer segments;
+	Py_ssize_t segmentCount;
+	Py_ssize_t i;
+
+	memset(&self->parent, 0, sizeof(self->parent));
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*y*:BufferWithSegments",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*s*:BufferWithSegments",
+#endif
+		kwlist, &self->parent, &segments)) {
+		return -1;
+	}
+
+	if (!PyBuffer_IsContiguous(&self->parent, 'C') || self->parent.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError, "data buffer should be contiguous and have a single dimension");
+		goto except;
+	}
+
+	if (!PyBuffer_IsContiguous(&segments, 'C') || segments.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError, "segments buffer should be contiguous and have a single dimension");
+		goto except;
+	}
+
+	if (segments.len % sizeof(BufferSegment)) {
+		PyErr_Format(PyExc_ValueError, "segments array size is not a multiple of %zu",
+			sizeof(BufferSegment));
+		goto except;
+	}
+
+	segmentCount = segments.len / sizeof(BufferSegment);
+
+	/* Validate segments data, as blindly trusting it could lead to arbitrary
+	memory access. */
+	for (i = 0; i < segmentCount; i++) {
+		BufferSegment* segment = &((BufferSegment*)(segments.buf))[i];
+
+		if (segment->offset + segment->length > (unsigned long long)self->parent.len) {
+			PyErr_SetString(PyExc_ValueError, "offset within segments array references memory outside buffer");
+			goto except;
+			return -1;
+		}
+	}
+
+	/* Make a copy of the segments data. It is cheap to do so and is a guard
+	   against caller changing offsets, which has security implications. */
+	self->segments = PyMem_Malloc(segments.len);
+	if (!self->segments) {
+		PyErr_NoMemory();
+		goto except;
+	}
+
+	memcpy(self->segments, segments.buf, segments.len);
+	PyBuffer_Release(&segments);
+
+	self->data = self->parent.buf;
+	self->dataSize = self->parent.len;
+	self->segmentCount = segmentCount;
+
+	return 0;
+
+except:
+	PyBuffer_Release(&self->parent);
+	PyBuffer_Release(&segments);
+	return -1;
+}
+
+/**
+ * Construct a BufferWithSegments from existing memory and offsets.
+ *
+ * Ownership of the backing memory and BufferSegments will be transferred to
+ * the created object and freed when the BufferWithSegments is destroyed.
+ */
+ZstdBufferWithSegments* BufferWithSegments_FromMemory(void* data, unsigned long long dataSize,
+	BufferSegment* segments, Py_ssize_t segmentsSize) {
+	ZstdBufferWithSegments* result = NULL;
+	Py_ssize_t i;
+
+	if (NULL == data) {
+		PyErr_SetString(PyExc_ValueError, "data is NULL");
+		return NULL;
+	}
+
+	if (NULL == segments) {
+		PyErr_SetString(PyExc_ValueError, "segments is NULL");
+		return NULL;
+	}
+
+	for (i = 0; i < segmentsSize; i++) {
+		BufferSegment* segment = &segments[i];
+
+		if (segment->offset + segment->length > dataSize) {
+			PyErr_SetString(PyExc_ValueError, "offset in segments overflows buffer size");
+			return NULL;
+		}
+	}
+
+	result = PyObject_New(ZstdBufferWithSegments, &ZstdBufferWithSegmentsType);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	result->useFree = 0;
+
+	memset(&result->parent, 0, sizeof(result->parent));
+	result->data = data;
+	result->dataSize = dataSize;
+	result->segments = segments;
+	result->segmentCount = segmentsSize;
+
+	return result;
+}
+
+static Py_ssize_t BufferWithSegments_length(ZstdBufferWithSegments* self) {
+	return self->segmentCount;
+}
+
+static ZstdBufferSegment* BufferWithSegments_item(ZstdBufferWithSegments* self, Py_ssize_t i) {
+	ZstdBufferSegment* result = NULL;
+
+	if (i < 0) {
+		PyErr_SetString(PyExc_IndexError, "offset must be non-negative");
+		return NULL;
+	}
+
+	if (i >= self->segmentCount) {
+		PyErr_Format(PyExc_IndexError, "offset must be less than %zd", self->segmentCount);
+		return NULL;
+	}
+
+	if (self->segments[i].length > PY_SSIZE_T_MAX) {
+		PyErr_Format(PyExc_ValueError,
+			"item at offset %zd is too large for this platform", i);
+		return NULL;
+	}
+
+	result = (ZstdBufferSegment*)PyObject_CallObject((PyObject*)&ZstdBufferSegmentType, NULL);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	result->parent = (PyObject*)self;
+	Py_INCREF(self);
+
+	result->data = (char*)self->data + self->segments[i].offset;
+	result->dataSize = (Py_ssize_t)self->segments[i].length;
+	result->offset = self->segments[i].offset;
+
+	return result;
+}
+
+#if PY_MAJOR_VERSION >= 3
+static int BufferWithSegments_getbuffer(ZstdBufferWithSegments* self, Py_buffer* view, int flags) {
+	if (self->dataSize > PY_SSIZE_T_MAX) {
+		view->obj = NULL;
+		PyErr_SetString(PyExc_BufferError, "buffer is too large for this platform");
+		return -1;
+	}
+
+	return PyBuffer_FillInfo(view, (PyObject*)self, self->data, (Py_ssize_t)self->dataSize, 1, flags);
+}
+#else
+static Py_ssize_t BufferWithSegments_getreadbuffer(ZstdBufferWithSegments* self, Py_ssize_t segment, void **ptrptr) {
+	if (segment != 0) {
+		PyErr_SetString(PyExc_ValueError, "segment number must be 0");
+		return -1;
+	}
+
+	if (self->dataSize > PY_SSIZE_T_MAX) {
+		PyErr_SetString(PyExc_ValueError, "buffer is too large for this platform");
+		return -1;
+	}
+
+	*ptrptr = self->data;
+	return (Py_ssize_t)self->dataSize;
+}
+
+static Py_ssize_t BufferWithSegments_getsegcount(ZstdBufferWithSegments* self, Py_ssize_t* len) {
+	if (len) {
+		*len = 1;
+	}
+
+	return 1;
+}
+#endif
+
+PyDoc_STRVAR(BufferWithSegments_tobytes__doc__,
+"Obtain a bytes instance for this buffer.\n"
+);
+
+static PyObject* BufferWithSegments_tobytes(ZstdBufferWithSegments* self) {
+	if (self->dataSize > PY_SSIZE_T_MAX) {
+		PyErr_SetString(PyExc_ValueError, "buffer is too large for this platform");
+		return NULL;
+	}
+
+	return PyBytes_FromStringAndSize(self->data, (Py_ssize_t)self->dataSize);
+}
+
+PyDoc_STRVAR(BufferWithSegments_segments__doc__,
+"Obtain a BufferSegments describing segments in this sintance.\n"
+);
+
+static ZstdBufferSegments* BufferWithSegments_segments(ZstdBufferWithSegments* self) {
+	ZstdBufferSegments* result = (ZstdBufferSegments*)PyObject_CallObject((PyObject*)&ZstdBufferSegmentsType, NULL);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	result->parent = (PyObject*)self;
+	Py_INCREF(self);
+	result->segments = self->segments;
+	result->segmentCount = self->segmentCount;
+
+	return result;
+}
+
+static PySequenceMethods BufferWithSegments_sq = {
+	(lenfunc)BufferWithSegments_length, /* sq_length */
+	0, /* sq_concat */
+	0, /* sq_repeat */
+	(ssizeargfunc)BufferWithSegments_item, /* sq_item */
+	0, /* sq_ass_item */
+	0, /* sq_contains */
+	0, /* sq_inplace_concat */
+	0 /* sq_inplace_repeat */
+};
+
+static PyBufferProcs BufferWithSegments_as_buffer = {
+#if PY_MAJOR_VERSION >= 3
+	(getbufferproc)BufferWithSegments_getbuffer, /* bf_getbuffer */
+	0 /* bf_releasebuffer */
+#else
+	(readbufferproc)BufferWithSegments_getreadbuffer, /* bf_getreadbuffer */
+	0, /* bf_getwritebuffer */
+	(segcountproc)BufferWithSegments_getsegcount, /* bf_getsegcount */
+	0 /* bf_getcharbuffer */
+#endif
+};
+
+static PyMethodDef BufferWithSegments_methods[] = {
+	{ "segments", (PyCFunction)BufferWithSegments_segments,
+	  METH_NOARGS, BufferWithSegments_segments__doc__ },
+	{ "tobytes", (PyCFunction)BufferWithSegments_tobytes,
+	  METH_NOARGS, BufferWithSegments_tobytes__doc__ },
+	{ NULL, NULL }
+};
+
+static PyMemberDef BufferWithSegments_members[] = {
+	{ "size", T_ULONGLONG, offsetof(ZstdBufferWithSegments, dataSize),
+	  READONLY, "total size of the buffer in bytes" },
+	{ NULL }
+};
+
+PyTypeObject ZstdBufferWithSegmentsType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.BufferWithSegments", /* tp_name */
+	sizeof(ZstdBufferWithSegments),/* tp_basicsize */
+	0,                         /* tp_itemsize */
+	(destructor)BufferWithSegments_dealloc, /* tp_dealloc */
+	0,                         /* tp_print */
+	0,                         /* tp_getattr */
+	0,                         /* tp_setattr */
+	0,                         /* tp_compare */
+	0,                         /* tp_repr */
+	0,                         /* tp_as_number */
+	&BufferWithSegments_sq,    /* tp_as_sequence */
+	0,                         /* tp_as_mapping */
+	0,                         /* tp_hash  */
+	0,                         /* tp_call */
+	0,                         /* tp_str */
+	0,                         /* tp_getattro */
+	0,                         /* tp_setattro */
+	&BufferWithSegments_as_buffer, /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT,        /* tp_flags */
+	BufferWithSegments__doc__, /* tp_doc */
+	0,                         /* tp_traverse */
+	0,                         /* tp_clear */
+	0,                         /* tp_richcompare */
+	0,                         /* tp_weaklistoffset */
+	0,                         /* tp_iter */
+	0,                         /* tp_iternext */
+	BufferWithSegments_methods, /* tp_methods */
+	BufferWithSegments_members, /* tp_members */
+	0,                         /* tp_getset */
+	0,                         /* tp_base */
+	0,                         /* tp_dict */
+	0,                         /* tp_descr_get */
+	0,                         /* tp_descr_set */
+	0,                         /* tp_dictoffset */
+	(initproc)BufferWithSegments_init, /* tp_init */
+	0,                         /* tp_alloc */
+	PyType_GenericNew,         /* tp_new */
+};
+
+PyDoc_STRVAR(BufferSegments__doc__,
+"BufferSegments - Represents segments/offsets within a BufferWithSegments\n"
+);
+
+static void BufferSegments_dealloc(ZstdBufferSegments* self) {
+	Py_CLEAR(self->parent);
+	PyObject_Del(self);
+}
+
+#if PY_MAJOR_VERSION >= 3
+static int BufferSegments_getbuffer(ZstdBufferSegments* self, Py_buffer* view, int flags) {
+	return PyBuffer_FillInfo(view, (PyObject*)self,
+		(void*)self->segments, self->segmentCount * sizeof(BufferSegment),
+		1, flags);
+}
+#else
+static Py_ssize_t BufferSegments_getreadbuffer(ZstdBufferSegments* self, Py_ssize_t segment, void **ptrptr) {
+	if (segment != 0) {
+		PyErr_SetString(PyExc_ValueError, "segment number must be 0");
+		return -1;
+	}
+
+	*ptrptr = (void*)self->segments;
+	return self->segmentCount * sizeof(BufferSegment);
+}
+
+static Py_ssize_t BufferSegments_getsegcount(ZstdBufferSegments* self, Py_ssize_t* len) {
+	if (len) {
+		*len = 1;
+	}
+
+	return 1;
+}
+#endif
+
+static PyBufferProcs BufferSegments_as_buffer = {
+#if PY_MAJOR_VERSION >= 3
+	(getbufferproc)BufferSegments_getbuffer,
+	0
+#else
+	(readbufferproc)BufferSegments_getreadbuffer,
+	0,
+	(segcountproc)BufferSegments_getsegcount,
+	0
+#endif
+};
+
+PyTypeObject ZstdBufferSegmentsType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.BufferSegments", /* tp_name */
+	sizeof(ZstdBufferSegments),/* tp_basicsize */
+	0,                         /* tp_itemsize */
+	(destructor)BufferSegments_dealloc, /* tp_dealloc */
+	0,                         /* tp_print */
+	0,                         /* tp_getattr */
+	0,                         /* tp_setattr */
+	0,                         /* tp_compare */
+	0,                         /* tp_repr */
+	0,                         /* tp_as_number */
+	0,                         /* tp_as_sequence */
+	0,                         /* tp_as_mapping */
+	0,                         /* tp_hash  */
+	0,                         /* tp_call */
+	0,                         /* tp_str */
+	0,                         /* tp_getattro */
+	0,                         /* tp_setattro */
+	&BufferSegments_as_buffer, /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT,        /* tp_flags */
+	BufferSegments__doc__,     /* tp_doc */
+	0,                         /* tp_traverse */
+	0,                         /* tp_clear */
+	0,                         /* tp_richcompare */
+	0,                         /* tp_weaklistoffset */
+	0,                         /* tp_iter */
+	0,                         /* tp_iternext */
+	0,                         /* tp_methods */
+	0,                         /* tp_members */
+	0,                         /* tp_getset */
+	0,                         /* tp_base */
+	0,                         /* tp_dict */
+	0,                         /* tp_descr_get */
+	0,                         /* tp_descr_set */
+	0,                         /* tp_dictoffset */
+	0,                         /* tp_init */
+	0,                         /* tp_alloc */
+	PyType_GenericNew,         /* tp_new */
+};
+
+PyDoc_STRVAR(BufferSegment__doc__,
+	"BufferSegment - Represents a segment within a BufferWithSegments\n"
+);
+
+static void BufferSegment_dealloc(ZstdBufferSegment* self) {
+	Py_CLEAR(self->parent);
+	PyObject_Del(self);
+}
+
+static Py_ssize_t BufferSegment_length(ZstdBufferSegment* self) {
+	return self->dataSize;
+}
+
+#if PY_MAJOR_VERSION >= 3
+static int BufferSegment_getbuffer(ZstdBufferSegment* self, Py_buffer* view, int flags) {
+	return PyBuffer_FillInfo(view, (PyObject*)self,
+		self->data, self->dataSize, 1, flags);
+}
+#else
+static Py_ssize_t BufferSegment_getreadbuffer(ZstdBufferSegment* self, Py_ssize_t segment, void **ptrptr) {
+	if (segment != 0) {
+		PyErr_SetString(PyExc_ValueError, "segment number must be 0");
+		return -1;
+	}
+
+	*ptrptr = self->data;
+	return self->dataSize;
+}
+
+static Py_ssize_t BufferSegment_getsegcount(ZstdBufferSegment* self, Py_ssize_t* len) {
+	if (len) {
+		*len = 1;
+	}
+
+	return 1;
+}
+#endif
+
+PyDoc_STRVAR(BufferSegment_tobytes__doc__,
+"Obtain a bytes instance for this segment.\n"
+);
+
+static PyObject* BufferSegment_tobytes(ZstdBufferSegment* self) {
+	return PyBytes_FromStringAndSize(self->data, self->dataSize);
+}
+
+static PySequenceMethods BufferSegment_sq = {
+	(lenfunc)BufferSegment_length, /* sq_length */
+	0, /* sq_concat */
+	0, /* sq_repeat */
+	0, /* sq_item */
+	0, /* sq_ass_item */
+	0, /* sq_contains */
+	0, /* sq_inplace_concat */
+	0 /* sq_inplace_repeat */
+};
+
+static PyBufferProcs BufferSegment_as_buffer = {
+#if PY_MAJOR_VERSION >= 3
+	(getbufferproc)BufferSegment_getbuffer,
+	0
+#else
+	(readbufferproc)BufferSegment_getreadbuffer,
+	0,
+	(segcountproc)BufferSegment_getsegcount,
+	0
+#endif
+};
+
+static PyMethodDef BufferSegment_methods[] = {
+	{ "tobytes", (PyCFunction)BufferSegment_tobytes,
+	  METH_NOARGS, BufferSegment_tobytes__doc__ },
+	{ NULL, NULL }
+};
+
+static PyMemberDef BufferSegment_members[] = {
+	{ "offset", T_ULONGLONG, offsetof(ZstdBufferSegment, offset), READONLY,
+	  "offset of segment within parent buffer" },
+	  { NULL }
+};
+
+PyTypeObject ZstdBufferSegmentType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.BufferSegment", /* tp_name */
+	sizeof(ZstdBufferSegment),/* tp_basicsize */
+	0,                         /* tp_itemsize */
+	(destructor)BufferSegment_dealloc, /* tp_dealloc */
+	0,                         /* tp_print */
+	0,                         /* tp_getattr */
+	0,                         /* tp_setattr */
+	0,                         /* tp_compare */
+	0,                         /* tp_repr */
+	0,                         /* tp_as_number */
+	&BufferSegment_sq,         /* tp_as_sequence */
+	0,                         /* tp_as_mapping */
+	0,                         /* tp_hash  */
+	0,                         /* tp_call */
+	0,                         /* tp_str */
+	0,                         /* tp_getattro */
+	0,                         /* tp_setattro */
+	&BufferSegment_as_buffer,  /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT,        /* tp_flags */
+	BufferSegment__doc__,      /* tp_doc */
+	0,                         /* tp_traverse */
+	0,                         /* tp_clear */
+	0,                         /* tp_richcompare */
+	0,                         /* tp_weaklistoffset */
+	0,                         /* tp_iter */
+	0,                         /* tp_iternext */
+	BufferSegment_methods,     /* tp_methods */
+	BufferSegment_members,     /* tp_members */
+	0,                         /* tp_getset */
+	0,                         /* tp_base */
+	0,                         /* tp_dict */
+	0,                         /* tp_descr_get */
+	0,                         /* tp_descr_set */
+	0,                         /* tp_dictoffset */
+	0,                         /* tp_init */
+	0,                         /* tp_alloc */
+	PyType_GenericNew,         /* tp_new */
+};
+
+PyDoc_STRVAR(BufferWithSegmentsCollection__doc__,
+"Represents a collection of BufferWithSegments.\n"
+);
+
+static void BufferWithSegmentsCollection_dealloc(ZstdBufferWithSegmentsCollection* self) {
+	Py_ssize_t i;
+
+	if (self->firstElements) {
+		PyMem_Free(self->firstElements);
+		self->firstElements = NULL;
+	}
+
+	if (self->buffers) {
+		for (i = 0; i < self->bufferCount; i++) {
+			Py_CLEAR(self->buffers[i]);
+		}
+
+		PyMem_Free(self->buffers);
+		self->buffers = NULL;
+	}
+
+	PyObject_Del(self);
+}
+
+static int BufferWithSegmentsCollection_init(ZstdBufferWithSegmentsCollection* self, PyObject* args) {
+	Py_ssize_t size;
+	Py_ssize_t i;
+	Py_ssize_t offset = 0;
+
+	size = PyTuple_Size(args);
+	if (-1 == size) {
+		return -1;
+	}
+
+	if (0 == size) {
+		PyErr_SetString(PyExc_ValueError, "must pass at least 1 argument");
+		return -1;
+	}
+
+	for (i = 0; i < size; i++) {
+		PyObject* item = PyTuple_GET_ITEM(args, i);
+		if (!PyObject_TypeCheck(item, &ZstdBufferWithSegmentsType)) {
+			PyErr_SetString(PyExc_TypeError, "arguments must be BufferWithSegments instances");
+			return -1;
+		}
+
+		if (0 == ((ZstdBufferWithSegments*)item)->segmentCount ||
+			0 == ((ZstdBufferWithSegments*)item)->dataSize) {
+			PyErr_SetString(PyExc_ValueError, "ZstdBufferWithSegments cannot be empty");
+			return -1;
+		}
+	}
+
+	self->buffers = PyMem_Malloc(size * sizeof(ZstdBufferWithSegments*));
+	if (NULL == self->buffers) {
+		PyErr_NoMemory();
+		return -1;
+	}
+
+	self->firstElements = PyMem_Malloc(size * sizeof(Py_ssize_t));
+	if (NULL == self->firstElements) {
+		PyMem_Free(self->buffers);
+		self->buffers = NULL;
+		PyErr_NoMemory();
+		return -1;
+	}
+
+	self->bufferCount = size;
+
+	for (i = 0; i < size; i++) {
+		ZstdBufferWithSegments* item = (ZstdBufferWithSegments*)PyTuple_GET_ITEM(args, i);
+
+		self->buffers[i] = item;
+		Py_INCREF(item);
+
+		if (i > 0) {
+			self->firstElements[i - 1] = offset;
+		}
+
+		offset += item->segmentCount;
+	}
+
+	self->firstElements[size - 1] = offset;
+
+	return 0;
+}
+
+static PyObject* BufferWithSegmentsCollection_size(ZstdBufferWithSegmentsCollection* self) {
+	Py_ssize_t i;
+	Py_ssize_t j;
+	unsigned long long size = 0;
+
+	for (i = 0; i < self->bufferCount; i++) {
+		for (j = 0; j < self->buffers[i]->segmentCount; j++) {
+			size += self->buffers[i]->segments[j].length;
+		}
+	}
+
+	return PyLong_FromUnsignedLongLong(size);
+}
+
+Py_ssize_t BufferWithSegmentsCollection_length(ZstdBufferWithSegmentsCollection* self) {
+	return self->firstElements[self->bufferCount - 1];
+}
+
+static ZstdBufferSegment* BufferWithSegmentsCollection_item(ZstdBufferWithSegmentsCollection* self, Py_ssize_t i) {
+	Py_ssize_t bufferOffset;
+
+	if (i < 0) {
+		PyErr_SetString(PyExc_IndexError, "offset must be non-negative");
+		return NULL;
+	}
+
+	if (i >= BufferWithSegmentsCollection_length(self)) {
+		PyErr_Format(PyExc_IndexError, "offset must be less than %zd",
+			BufferWithSegmentsCollection_length(self));
+		return NULL;
+	}
+
+	for (bufferOffset = 0; bufferOffset < self->bufferCount; bufferOffset++) {
+		Py_ssize_t offset = 0;
+
+		if (i < self->firstElements[bufferOffset]) {
+			if (bufferOffset > 0) {
+				offset = self->firstElements[bufferOffset - 1];
+			}
+
+			return BufferWithSegments_item(self->buffers[bufferOffset], i - offset);
+		}
+	}
+
+	PyErr_SetString(ZstdError, "error resolving segment; this should not happen");
+	return NULL;
+}
+
+static PySequenceMethods BufferWithSegmentsCollection_sq = {
+	(lenfunc)BufferWithSegmentsCollection_length, /* sq_length */
+	0, /* sq_concat */
+	0, /* sq_repeat */
+	(ssizeargfunc)BufferWithSegmentsCollection_item, /* sq_item */
+	0, /* sq_ass_item */
+	0, /* sq_contains */
+	0, /* sq_inplace_concat */
+	0 /* sq_inplace_repeat */
+};
+
+static PyMethodDef BufferWithSegmentsCollection_methods[] = {
+	{ "size", (PyCFunction)BufferWithSegmentsCollection_size,
+	  METH_NOARGS, PyDoc_STR("total size in bytes of all segments") },
+	{ NULL, NULL }
+};
+
+PyTypeObject ZstdBufferWithSegmentsCollectionType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.BufferWithSegmentsCollection", /* tp_name */
+	sizeof(ZstdBufferWithSegmentsCollection),/* tp_basicsize */
+	0,                         /* tp_itemsize */
+	(destructor)BufferWithSegmentsCollection_dealloc, /* tp_dealloc */
+	0,                         /* tp_print */
+	0,                         /* tp_getattr */
+	0,                         /* tp_setattr */
+	0,                         /* tp_compare */
+	0,                         /* tp_repr */
+	0,                         /* tp_as_number */
+	&BufferWithSegmentsCollection_sq, /* tp_as_sequence */
+	0,                         /* tp_as_mapping */
+	0,                         /* tp_hash  */
+	0,                         /* tp_call */
+	0,                         /* tp_str */
+	0,                         /* tp_getattro */
+	0,                         /* tp_setattro */
+	0,                         /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT,        /* tp_flags */
+	BufferWithSegmentsCollection__doc__, /* tp_doc */
+	0,                         /* tp_traverse */
+	0,                         /* tp_clear */
+	0,                         /* tp_richcompare */
+	0,                         /* tp_weaklistoffset */
+	/* TODO implement iterator for performance. */
+	0,                         /* tp_iter */
+	0,                         /* tp_iternext */
+	BufferWithSegmentsCollection_methods, /* tp_methods */
+	0,                         /* tp_members */
+	0,                         /* tp_getset */
+	0,                         /* tp_base */
+	0,                         /* tp_dict */
+	0,                         /* tp_descr_get */
+	0,                         /* tp_descr_set */
+	0,                         /* tp_dictoffset */
+	(initproc)BufferWithSegmentsCollection_init, /* tp_init */
+	0,                         /* tp_alloc */
+	PyType_GenericNew,         /* tp_new */
+};
+
+void bufferutil_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdBufferWithSegmentsType) = &PyType_Type;
+	if (PyType_Ready(&ZstdBufferWithSegmentsType) < 0) {
+		return;
+	}
+
+	Py_INCREF(&ZstdBufferWithSegmentsType);
+	PyModule_AddObject(mod, "BufferWithSegments", (PyObject*)&ZstdBufferWithSegmentsType);
+
+	Py_TYPE(&ZstdBufferSegmentsType) = &PyType_Type;
+	if (PyType_Ready(&ZstdBufferSegmentsType) < 0) {
+		return;
+	}
+
+	Py_INCREF(&ZstdBufferSegmentsType);
+	PyModule_AddObject(mod, "BufferSegments", (PyObject*)&ZstdBufferSegmentsType);
+
+	Py_TYPE(&ZstdBufferSegmentType) = &PyType_Type;
+	if (PyType_Ready(&ZstdBufferSegmentType) < 0) {
+		return;
+	}
+
+	Py_INCREF(&ZstdBufferSegmentType);
+	PyModule_AddObject(mod, "BufferSegment", (PyObject*)&ZstdBufferSegmentType);
+
+	Py_TYPE(&ZstdBufferWithSegmentsCollectionType) = &PyType_Type;
+	if (PyType_Ready(&ZstdBufferWithSegmentsCollectionType) < 0) {
+		return;
+	}
+
+	Py_INCREF(&ZstdBufferWithSegmentsCollectionType);
+	PyModule_AddObject(mod, "BufferWithSegmentsCollection", (PyObject*)&ZstdBufferWithSegmentsCollectionType);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressionchunker.c b/contrib/python/zstandard/py2/c-ext/compressionchunker.c
new file mode 100644
index 00000000000..6677ebe59f1
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressionchunker.c
@@ -0,0 +1,360 @@
+/**
+* Copyright (c) 2018-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(ZstdCompressionChunkerIterator__doc__,
+	"Iterator of output chunks from ZstdCompressionChunker.\n"
+);
+
+static void ZstdCompressionChunkerIterator_dealloc(ZstdCompressionChunkerIterator* self) {
+	Py_XDECREF(self->chunker);
+
+	PyObject_Del(self);
+}
+
+static PyObject* ZstdCompressionChunkerIterator_iter(PyObject* self) {
+	Py_INCREF(self);
+	return self;
+}
+
+static PyObject* ZstdCompressionChunkerIterator_iternext(ZstdCompressionChunkerIterator* self) {
+	size_t zresult;
+	PyObject* chunk;
+	ZstdCompressionChunker* chunker = self->chunker;
+	ZSTD_EndDirective zFlushMode;
+
+	if (self->mode != compressionchunker_mode_normal && chunker->input.pos != chunker->input.size) {
+		PyErr_SetString(ZstdError, "input should have been fully consumed before calling flush() or finish()");
+		return NULL;
+	}
+
+	if (chunker->finished) {
+		return NULL;
+	}
+
+	/* If we have data left in the input, consume it. */
+	while (chunker->input.pos < chunker->input.size) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(chunker->compressor->cctx, &chunker->output,
+			&chunker->input, ZSTD_e_continue);
+		Py_END_ALLOW_THREADS
+
+		/* Input is fully consumed. */
+		if (chunker->input.pos == chunker->input.size) {
+			chunker->input.src = NULL;
+			chunker->input.pos = 0;
+			chunker->input.size = 0;
+			PyBuffer_Release(&chunker->inBuffer);
+		}
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+			return NULL;
+		}
+
+		/* If it produced a full output chunk, emit it. */
+		if (chunker->output.pos == chunker->output.size) {
+			chunk = PyBytes_FromStringAndSize(chunker->output.dst, chunker->output.pos);
+			if (!chunk) {
+				return NULL;
+			}
+
+			chunker->output.pos = 0;
+
+			return chunk;
+		}
+
+		/* Else continue to compress available input data. */
+	}
+
+	/* We also need this here for the special case of an empty input buffer. */
+	if (chunker->input.pos == chunker->input.size) {
+		chunker->input.src = NULL;
+		chunker->input.pos = 0;
+		chunker->input.size = 0;
+		PyBuffer_Release(&chunker->inBuffer);
+	}
+
+	/* No more input data. A partial chunk may be in chunker->output.
+	 * If we're in normal compression mode, we're done. Otherwise if we're in
+	 * flush or finish mode, we need to emit what data remains.
+	 */
+	if (self->mode == compressionchunker_mode_normal) {
+		/* We don't need to set StopIteration. */
+		return NULL;
+	}
+
+	if (self->mode == compressionchunker_mode_flush) {
+		zFlushMode = ZSTD_e_flush;
+	}
+	else if (self->mode == compressionchunker_mode_finish) {
+		zFlushMode = ZSTD_e_end;
+	}
+	else {
+		PyErr_SetString(ZstdError, "unhandled compression mode; this should never happen");
+		return NULL;
+	}
+
+	Py_BEGIN_ALLOW_THREADS
+	zresult = ZSTD_compressStream2(chunker->compressor->cctx, &chunker->output,
+		&chunker->input, zFlushMode);
+	Py_END_ALLOW_THREADS
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "zstd compress error: %s",
+			ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	if (!zresult && chunker->output.pos == 0) {
+		return NULL;
+	}
+
+	chunk = PyBytes_FromStringAndSize(chunker->output.dst, chunker->output.pos);
+	if (!chunk) {
+		return NULL;
+	}
+
+	chunker->output.pos = 0;
+
+	if (!zresult && self->mode == compressionchunker_mode_finish) {
+		chunker->finished = 1;
+	}
+
+	return chunk;
+}
+
+PyTypeObject ZstdCompressionChunkerIteratorType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressionChunkerIterator", /* tp_name */
+	sizeof(ZstdCompressionChunkerIterator), /* tp_basicsize */
+	0,                               /* tp_itemsize */
+	(destructor)ZstdCompressionChunkerIterator_dealloc, /* tp_dealloc */
+	0,                               /* tp_print */
+	0,                               /* tp_getattr */
+	0,                               /* tp_setattr */
+	0,                               /* tp_compare */
+	0,                               /* tp_repr */
+	0,                               /* tp_as_number */
+	0,                               /* tp_as_sequence */
+	0,                               /* tp_as_mapping */
+	0,                               /* tp_hash */
+	0,                               /* tp_call */
+	0,                               /* tp_str */
+	0,                               /* tp_getattro */
+	0,                               /* tp_setattro */
+	0,                               /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressionChunkerIterator__doc__, /* tp_doc */
+	0,                               /* tp_traverse */
+	0,                               /* tp_clear */
+	0,                               /* tp_richcompare */
+	0,                               /* tp_weaklistoffset */
+	ZstdCompressionChunkerIterator_iter, /* tp_iter */
+	(iternextfunc)ZstdCompressionChunkerIterator_iternext, /* tp_iternext */
+	0,                               /* tp_methods */
+	0,                               /* tp_members */
+	0,                               /* tp_getset */
+	0,                               /* tp_base */
+	0,                               /* tp_dict */
+	0,                               /* tp_descr_get */
+	0,                               /* tp_descr_set */
+	0,                               /* tp_dictoffset */
+	0,                               /* tp_init */
+	0,                               /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+PyDoc_STRVAR(ZstdCompressionChunker__doc__,
+	"Compress chunks iteratively into exact chunk sizes.\n"
+);
+
+static void ZstdCompressionChunker_dealloc(ZstdCompressionChunker* self) {
+	PyBuffer_Release(&self->inBuffer);
+	self->input.src = NULL;
+
+	PyMem_Free(self->output.dst);
+	self->output.dst = NULL;
+
+	Py_XDECREF(self->compressor);
+
+	PyObject_Del(self);
+}
+
+static ZstdCompressionChunkerIterator* ZstdCompressionChunker_compress(ZstdCompressionChunker* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	ZstdCompressionChunkerIterator* result;
+
+	if (self->finished) {
+		PyErr_SetString(ZstdError, "cannot call compress() after compression finished");
+		return NULL;
+	}
+
+	if (self->inBuffer.obj) {
+		PyErr_SetString(ZstdError,
+			"cannot perform operation before consuming output from previous operation");
+		return NULL;
+	}
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:compress",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:compress",
+#endif
+		kwlist, &self->inBuffer)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&self->inBuffer, 'C') || self->inBuffer.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		PyBuffer_Release(&self->inBuffer);
+		return NULL;
+	}
+
+	result = (ZstdCompressionChunkerIterator*)PyObject_CallObject((PyObject*)&ZstdCompressionChunkerIteratorType, NULL);
+	if (!result) {
+		PyBuffer_Release(&self->inBuffer);
+		return NULL;
+	}
+
+	self->input.src = self->inBuffer.buf;
+	self->input.size = self->inBuffer.len;
+	self->input.pos = 0;
+
+	result->chunker = self;
+	Py_INCREF(result->chunker);
+
+	result->mode = compressionchunker_mode_normal;
+
+	return result;
+}
+
+static ZstdCompressionChunkerIterator* ZstdCompressionChunker_finish(ZstdCompressionChunker* self) {
+	ZstdCompressionChunkerIterator* result;
+
+	if (self->finished) {
+		PyErr_SetString(ZstdError, "cannot call finish() after compression finished");
+		return NULL;
+	}
+
+	if (self->inBuffer.obj) {
+		PyErr_SetString(ZstdError,
+			"cannot call finish() before consuming output from previous operation");
+		return NULL;
+	}
+
+	result = (ZstdCompressionChunkerIterator*)PyObject_CallObject((PyObject*)&ZstdCompressionChunkerIteratorType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	result->chunker = self;
+	Py_INCREF(result->chunker);
+
+	result->mode = compressionchunker_mode_finish;
+
+	return result;
+}
+
+static ZstdCompressionChunkerIterator* ZstdCompressionChunker_flush(ZstdCompressionChunker* self, PyObject* args, PyObject* kwargs) {
+	ZstdCompressionChunkerIterator* result;
+
+	if (self->finished) {
+		PyErr_SetString(ZstdError, "cannot call flush() after compression finished");
+		return NULL;
+	}
+
+	if (self->inBuffer.obj) {
+		PyErr_SetString(ZstdError,
+			"cannot call flush() before consuming output from previous operation");
+		return NULL;
+	}
+
+	result = (ZstdCompressionChunkerIterator*)PyObject_CallObject((PyObject*)&ZstdCompressionChunkerIteratorType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	result->chunker = self;
+	Py_INCREF(result->chunker);
+
+	result->mode = compressionchunker_mode_flush;
+
+	return result;
+}
+
+static PyMethodDef ZstdCompressionChunker_methods[] = {
+	{ "compress", (PyCFunction)ZstdCompressionChunker_compress, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("compress data") },
+	{ "finish", (PyCFunction)ZstdCompressionChunker_finish, METH_NOARGS,
+	PyDoc_STR("finish compression operation") },
+	{ "flush", (PyCFunction)ZstdCompressionChunker_flush, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("finish compression operation") },
+	{ NULL, NULL }
+};
+
+PyTypeObject ZstdCompressionChunkerType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressionChunkerType",  /* tp_name */
+	sizeof(ZstdCompressionChunker),     /* tp_basicsize */
+	0,                                  /* tp_itemsize */
+	(destructor)ZstdCompressionChunker_dealloc, /* tp_dealloc */
+	0,                                  /* tp_print */
+	0,                                  /* tp_getattr */
+	0,                                  /* tp_setattr */
+	0,                                  /* tp_compare */
+	0,                                  /* tp_repr */
+	0,                                  /* tp_as_number */
+	0,                                  /* tp_as_sequence */
+	0,                                  /* tp_as_mapping */
+	0,                                  /* tp_hash */
+	0,                                  /* tp_call */
+	0,                                  /* tp_str */
+	0,                                  /* tp_getattro */
+	0,                                  /* tp_setattro */
+	0,                                  /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressionChunker__doc__,      /* tp_doc */
+	0,                                  /* tp_traverse */
+	0,                                  /* tp_clear */
+	0,                                  /* tp_richcompare */
+	0,                                  /* tp_weaklistoffset */
+	0,                                  /* tp_iter */
+	0,                                  /* tp_iternext */
+	ZstdCompressionChunker_methods,     /* tp_methods */
+	0,                                  /* tp_members */
+	0,                                  /* tp_getset */
+	0,                                  /* tp_base */
+	0,                                  /* tp_dict */
+	0,                                  /* tp_descr_get */
+	0,                                  /* tp_descr_set */
+	0,                                  /* tp_dictoffset */
+	0,                                  /* tp_init */
+	0,                                  /* tp_alloc */
+	PyType_GenericNew,                  /* tp_new */
+};
+
+void compressionchunker_module_init(PyObject* module) {
+	Py_TYPE(&ZstdCompressionChunkerIteratorType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionChunkerIteratorType) < 0) {
+		return;
+	}
+
+	Py_TYPE(&ZstdCompressionChunkerType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionChunkerType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressiondict.c b/contrib/python/zstandard/py2/c-ext/compressiondict.c
new file mode 100644
index 00000000000..1379861648e
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressiondict.c
@@ -0,0 +1,411 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+ZstdCompressionDict* train_dictionary(PyObject* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"dict_size",
+		"samples",
+		"k",
+		"d",
+		"notifications",
+		"dict_id",
+		"level",
+		"steps",
+		"threads",
+		NULL
+	};
+
+	size_t capacity;
+	PyObject* samples;
+	unsigned k = 0;
+	unsigned d = 0;
+	unsigned notifications = 0;
+	unsigned dictID = 0;
+	int level = 0;
+	unsigned steps = 0;
+	int threads = 0;
+	ZDICT_cover_params_t params;
+	Py_ssize_t samplesLen;
+	Py_ssize_t i;
+	size_t samplesSize = 0;
+	void* sampleBuffer = NULL;
+	size_t* sampleSizes = NULL;
+	void* sampleOffset;
+	Py_ssize_t sampleSize;
+	void* dict = NULL;
+	size_t zresult;
+	ZstdCompressionDict* result = NULL;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "nO!|IIIIiIi:train_dictionary",
+		kwlist, &capacity, &PyList_Type, &samples,
+		&k, &d, &notifications, &dictID, &level, &steps, &threads)) {
+		return NULL;
+	}
+
+	if (threads < 0) {
+		threads = cpu_count();
+	}
+
+	memset(&params, 0, sizeof(params));
+	params.k = k;
+	params.d = d;
+	params.steps = steps;
+	params.nbThreads = threads;
+	params.zParams.notificationLevel = notifications;
+	params.zParams.dictID = dictID;
+	params.zParams.compressionLevel = level;
+
+	/* Figure out total size of input samples. */
+	samplesLen = PyList_Size(samples);
+	for (i = 0; i < samplesLen; i++) {
+		PyObject* sampleItem = PyList_GET_ITEM(samples, i);
+
+		if (!PyBytes_Check(sampleItem)) {
+			PyErr_SetString(PyExc_ValueError, "samples must be bytes");
+			return NULL;
+		}
+		samplesSize += PyBytes_GET_SIZE(sampleItem);
+	}
+
+	sampleBuffer = PyMem_Malloc(samplesSize);
+	if (!sampleBuffer) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+
+	sampleSizes = PyMem_Malloc(samplesLen * sizeof(size_t));
+	if (!sampleSizes) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+
+	sampleOffset = sampleBuffer;
+	for (i = 0; i < samplesLen; i++) {
+		PyObject* sampleItem = PyList_GET_ITEM(samples, i);
+		sampleSize = PyBytes_GET_SIZE(sampleItem);
+		sampleSizes[i] = sampleSize;
+		memcpy(sampleOffset, PyBytes_AS_STRING(sampleItem), sampleSize);
+		sampleOffset = (char*)sampleOffset + sampleSize;
+	}
+
+	dict = PyMem_Malloc(capacity);
+	if (!dict) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+
+	Py_BEGIN_ALLOW_THREADS
+	/* No parameters uses the default function, which will use default params
+	   and call ZDICT_optimizeTrainFromBuffer_cover under the hood. */
+	if (!params.k && !params.d && !params.zParams.compressionLevel
+		&& !params.zParams.notificationLevel && !params.zParams.dictID) {
+		zresult = ZDICT_trainFromBuffer(dict, capacity, sampleBuffer,
+			sampleSizes, (unsigned)samplesLen);
+	}
+	/* Use optimize mode if user controlled steps or threads explicitly. */
+	else if (params.steps || params.nbThreads) {
+		zresult = ZDICT_optimizeTrainFromBuffer_cover(dict, capacity,
+			sampleBuffer, sampleSizes, (unsigned)samplesLen, &params);
+	}
+	/* Non-optimize mode with explicit control. */
+	else {
+		zresult = ZDICT_trainFromBuffer_cover(dict, capacity,
+			sampleBuffer, sampleSizes, (unsigned)samplesLen, params);
+	}
+	Py_END_ALLOW_THREADS
+
+	if (ZDICT_isError(zresult)) {
+		PyMem_Free(dict);
+		PyErr_Format(ZstdError, "cannot train dict: %s", ZDICT_getErrorName(zresult));
+		goto finally;
+	}
+
+	result = PyObject_New(ZstdCompressionDict, &ZstdCompressionDictType);
+	if (!result) {
+		PyMem_Free(dict);
+		goto finally;
+	}
+
+	result->dictData = dict;
+	result->dictSize = zresult;
+	result->dictType = ZSTD_dct_fullDict;
+	result->d = params.d;
+	result->k = params.k;
+	result->cdict = NULL;
+	result->ddict = NULL;
+
+finally:
+	PyMem_Free(sampleBuffer);
+	PyMem_Free(sampleSizes);
+
+	return result;
+}
+
+int ensure_ddict(ZstdCompressionDict* dict) {
+	if (dict->ddict) {
+		return 0;
+	}
+
+	Py_BEGIN_ALLOW_THREADS
+	dict->ddict = ZSTD_createDDict_advanced(dict->dictData, dict->dictSize,
+		ZSTD_dlm_byRef, dict->dictType, ZSTD_defaultCMem);
+	Py_END_ALLOW_THREADS
+	if (!dict->ddict) {
+		PyErr_SetString(ZstdError, "could not create decompression dict");
+		return 1;
+	}
+
+	return 0;
+}
+
+PyDoc_STRVAR(ZstdCompressionDict__doc__,
+"ZstdCompressionDict(data) - Represents a computed compression dictionary\n"
+"\n"
+"This type holds the results of a computed Zstandard compression dictionary.\n"
+"Instances are obtained by calling ``train_dictionary()`` or by passing\n"
+"bytes obtained from another source into the constructor.\n"
+);
+
+static int ZstdCompressionDict_init(ZstdCompressionDict* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		"dict_type",
+		NULL
+	};
+
+	int result = -1;
+	Py_buffer source;
+	unsigned dictType = ZSTD_dct_auto;
+
+	self->dictData = NULL;
+	self->dictSize = 0;
+	self->cdict = NULL;
+	self->ddict = NULL;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|I:ZstdCompressionDict",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*|I:ZstdCompressionDict",
+#endif
+		kwlist, &source, &dictType)) {
+		return -1;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	if (dictType != ZSTD_dct_auto && dictType != ZSTD_dct_rawContent
+		&& dictType != ZSTD_dct_fullDict) {
+		PyErr_Format(PyExc_ValueError,
+			"invalid dictionary load mode: %d; must use DICT_TYPE_* constants",
+			dictType);
+		goto finally;
+	}
+
+	self->dictType = dictType;
+
+	self->dictData = PyMem_Malloc(source.len);
+	if (!self->dictData) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+
+	memcpy(self->dictData, source.buf, source.len);
+	self->dictSize = source.len;
+
+	result = 0;
+
+finally:
+	PyBuffer_Release(&source);
+	return result;
+}
+
+static void ZstdCompressionDict_dealloc(ZstdCompressionDict* self) {
+	if (self->cdict) {
+		ZSTD_freeCDict(self->cdict);
+		self->cdict = NULL;
+	}
+
+	if (self->ddict) {
+		ZSTD_freeDDict(self->ddict);
+		self->ddict = NULL;
+	}
+
+	if (self->dictData) {
+		PyMem_Free(self->dictData);
+		self->dictData = NULL;
+	}
+
+	PyObject_Del(self);
+}
+
+PyDoc_STRVAR(ZstdCompressionDict_precompute_compress__doc__,
+"Precompute a dictionary so it can be used by multiple compressors.\n"
+);
+
+static PyObject* ZstdCompressionDict_precompute_compress(ZstdCompressionDict* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"level",
+		"compression_params",
+		NULL
+	};
+
+	int level = 0;
+	ZstdCompressionParametersObject* compressionParams = NULL;
+	ZSTD_compressionParameters cParams;
+	size_t zresult;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|iO!:precompute_compress", kwlist,
+		&level, &ZstdCompressionParametersType, &compressionParams)) {
+		return NULL;
+	}
+
+	if (level && compressionParams) {
+		PyErr_SetString(PyExc_ValueError,
+			"must only specify one of level or compression_params");
+		return NULL;
+	}
+
+	if (!level && !compressionParams) {
+		PyErr_SetString(PyExc_ValueError,
+			"must specify one of level or compression_params");
+		return NULL;
+	}
+
+	if (self->cdict) {
+		zresult = ZSTD_freeCDict(self->cdict);
+		self->cdict = NULL;
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "unable to free CDict: %s",
+				ZSTD_getErrorName(zresult));
+			return NULL;
+		}
+	}
+
+	if (level) {
+		cParams = ZSTD_getCParams(level, 0, self->dictSize);
+	}
+	else {
+		if (to_cparams(compressionParams, &cParams)) {
+			return NULL;
+		}
+	}
+
+	assert(!self->cdict);
+	self->cdict = ZSTD_createCDict_advanced(self->dictData, self->dictSize,
+		ZSTD_dlm_byRef, self->dictType, cParams, ZSTD_defaultCMem);
+
+	if (!self->cdict) {
+		PyErr_SetString(ZstdError, "unable to precompute dictionary");
+		return NULL;
+	}
+
+	Py_RETURN_NONE;
+}
+
+static PyObject* ZstdCompressionDict_dict_id(ZstdCompressionDict* self) {
+	unsigned dictID = ZDICT_getDictID(self->dictData, self->dictSize);
+
+	return PyLong_FromLong(dictID);
+}
+
+static PyObject* ZstdCompressionDict_as_bytes(ZstdCompressionDict* self) {
+	return PyBytes_FromStringAndSize(self->dictData, self->dictSize);
+}
+
+static PyMethodDef ZstdCompressionDict_methods[] = {
+	{ "dict_id", (PyCFunction)ZstdCompressionDict_dict_id, METH_NOARGS,
+	PyDoc_STR("dict_id() -- obtain the numeric dictionary ID") },
+	{ "as_bytes", (PyCFunction)ZstdCompressionDict_as_bytes, METH_NOARGS,
+	PyDoc_STR("as_bytes() -- obtain the raw bytes constituting the dictionary data") },
+	{ "precompute_compress", (PyCFunction)ZstdCompressionDict_precompute_compress,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressionDict_precompute_compress__doc__ },
+	{ NULL, NULL }
+};
+
+static PyMemberDef ZstdCompressionDict_members[] = {
+	{ "k", T_UINT, offsetof(ZstdCompressionDict, k), READONLY,
+	  "segment size" },
+	{ "d", T_UINT, offsetof(ZstdCompressionDict, d), READONLY,
+	  "dmer size" },
+	{ NULL }
+};
+
+static Py_ssize_t ZstdCompressionDict_length(ZstdCompressionDict* self) {
+	return self->dictSize;
+}
+
+static PySequenceMethods ZstdCompressionDict_sq = {
+	(lenfunc)ZstdCompressionDict_length, /* sq_length */
+	0,                                   /* sq_concat */
+	0,                                   /* sq_repeat */
+	0,                                   /* sq_item */
+	0,                                   /* sq_ass_item */
+	0,                                   /* sq_contains */
+	0,                                   /* sq_inplace_concat */
+	0                                    /* sq_inplace_repeat */
+};
+
+PyTypeObject ZstdCompressionDictType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressionDict",     /* tp_name */
+	sizeof(ZstdCompressionDict),    /* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)ZstdCompressionDict_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	&ZstdCompressionDict_sq,        /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressionDict__doc__,     /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	ZstdCompressionDict_methods,    /* tp_methods */
+	ZstdCompressionDict_members,    /* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	(initproc)ZstdCompressionDict_init, /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void compressiondict_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdCompressionDictType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionDictType) < 0) {
+		return;
+	}
+
+	Py_INCREF((PyObject*)&ZstdCompressionDictType);
+	PyModule_AddObject(mod, "ZstdCompressionDict",
+		(PyObject*)&ZstdCompressionDictType);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressionparams.c b/contrib/python/zstandard/py2/c-ext/compressionparams.c
new file mode 100644
index 00000000000..e5e8c55fea2
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressionparams.c
@@ -0,0 +1,572 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+int set_parameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int value) {
+	size_t zresult = ZSTD_CCtxParams_setParameter(params, param, value);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "unable to set compression context parameter: %s",
+			ZSTD_getErrorName(zresult));
+		return 1;
+	}
+
+	return 0;
+}
+
+#define TRY_SET_PARAMETER(params, param, value) if (set_parameter(params, param, value)) return -1;
+
+#define TRY_COPY_PARAMETER(source, dest, param) { \
+	int result; \
+	size_t zresult = ZSTD_CCtxParams_getParameter(source, param, &result); \
+	if (ZSTD_isError(zresult)) { \
+		return 1; \
+	} \
+	zresult = ZSTD_CCtxParams_setParameter(dest, param, result); \
+	if (ZSTD_isError(zresult)) { \
+		return 1; \
+	} \
+}
+
+int set_parameters(ZSTD_CCtx_params* params, ZstdCompressionParametersObject* obj) {
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_nbWorkers);
+
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_format);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_compressionLevel);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_windowLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_hashLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_chainLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_searchLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_minMatch);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_targetLength);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_strategy);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_contentSizeFlag);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_checksumFlag);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_dictIDFlag);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_jobSize);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_overlapLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_forceMaxWindow);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_enableLongDistanceMatching);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmHashLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmMinMatch);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmBucketSizeLog);
+	TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmHashRateLog);
+
+	return 0;
+}
+
+int reset_params(ZstdCompressionParametersObject* params) {
+	if (params->params) {
+		ZSTD_CCtxParams_reset(params->params);
+	}
+	else {
+		params->params = ZSTD_createCCtxParams();
+		if (!params->params) {
+			PyErr_NoMemory();
+			return 1;
+		}
+	}
+
+	return set_parameters(params->params, params);
+}
+
+#define TRY_GET_PARAMETER(params, param, value) { \
+    size_t zresult = ZSTD_CCtxParams_getParameter(params, param, value); \
+    if (ZSTD_isError(zresult)) { \
+        PyErr_Format(ZstdError, "unable to retrieve parameter: %s", ZSTD_getErrorName(zresult)); \
+        return 1; \
+    } \
+}
+
+int to_cparams(ZstdCompressionParametersObject* params, ZSTD_compressionParameters* cparams) {
+	int value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_windowLog, &value);
+	cparams->windowLog = value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_chainLog, &value);
+	cparams->chainLog = value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_hashLog, &value);
+	cparams->hashLog = value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_searchLog, &value);
+	cparams->searchLog = value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_minMatch, &value);
+	cparams->minMatch = value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_targetLength, &value);
+	cparams->targetLength = value;
+
+	TRY_GET_PARAMETER(params->params, ZSTD_c_strategy, &value);
+	cparams->strategy = value;
+
+	return 0;
+}
+
+static int ZstdCompressionParameters_init(ZstdCompressionParametersObject* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"format",
+		"compression_level",
+		"window_log",
+		"hash_log",
+		"chain_log",
+		"search_log",
+		"min_match",
+		"target_length",
+		"compression_strategy",
+		"strategy",
+		"write_content_size",
+		"write_checksum",
+		"write_dict_id",
+		"job_size",
+		"overlap_log",
+		"overlap_size_log",
+		"force_max_window",
+		"enable_ldm",
+		"ldm_hash_log",
+		"ldm_min_match",
+		"ldm_bucket_size_log",
+		"ldm_hash_rate_log",
+		"ldm_hash_every_log",
+		"threads",
+		NULL
+	};
+
+	int format = 0;
+	int compressionLevel = 0;
+	int windowLog = 0;
+	int hashLog = 0;
+	int chainLog = 0;
+	int searchLog = 0;
+	int minMatch = 0;
+	int targetLength = 0;
+	int compressionStrategy = -1;
+	int strategy = -1;
+	int contentSizeFlag = 1;
+	int checksumFlag = 0;
+	int dictIDFlag = 0;
+	int jobSize = 0;
+	int overlapLog = -1;
+	int overlapSizeLog = -1;
+	int forceMaxWindow = 0;
+	int enableLDM = 0;
+	int ldmHashLog = 0;
+	int ldmMinMatch = 0;
+	int ldmBucketSizeLog = 0;
+	int ldmHashRateLog = -1;
+	int ldmHashEveryLog = -1;
+	int threads = 0;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs,
+		"|iiiiiiiiiiiiiiiiiiiiiiii:CompressionParameters",
+		kwlist, &format, &compressionLevel, &windowLog, &hashLog, &chainLog,
+		&searchLog, &minMatch, &targetLength, &compressionStrategy, &strategy,
+		&contentSizeFlag, &checksumFlag, &dictIDFlag, &jobSize, &overlapLog,
+		&overlapSizeLog, &forceMaxWindow, &enableLDM, &ldmHashLog, &ldmMinMatch,
+		&ldmBucketSizeLog, &ldmHashRateLog, &ldmHashEveryLog, &threads)) {
+		return -1;
+	}
+
+	if (reset_params(self)) {
+		return -1;
+	}
+
+	if (threads < 0) {
+		threads = cpu_count();
+	}
+
+	/* We need to set ZSTD_c_nbWorkers before ZSTD_c_jobSize and ZSTD_c_overlapLog
+	 * because setting ZSTD_c_nbWorkers resets the other parameters. */
+	TRY_SET_PARAMETER(self->params, ZSTD_c_nbWorkers, threads);
+
+	TRY_SET_PARAMETER(self->params, ZSTD_c_format, format);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_compressionLevel, compressionLevel);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_windowLog, windowLog);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_hashLog, hashLog);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_chainLog, chainLog);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_searchLog, searchLog);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_minMatch, minMatch);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_targetLength, targetLength);
+
+	if (compressionStrategy != -1 && strategy != -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot specify both compression_strategy and strategy");
+		return -1;
+    }
+
+	if (compressionStrategy != -1) {
+		strategy = compressionStrategy;
+	}
+	else if (strategy == -1) {
+		strategy = 0;
+	}
+
+	TRY_SET_PARAMETER(self->params, ZSTD_c_strategy, strategy);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_contentSizeFlag, contentSizeFlag);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_checksumFlag, checksumFlag);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_dictIDFlag, dictIDFlag);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_jobSize, jobSize);
+
+	if (overlapLog != -1 && overlapSizeLog != -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot specify both overlap_log and overlap_size_log");
+		return -1;
+	}
+
+	if (overlapSizeLog != -1) {
+		overlapLog = overlapSizeLog;
+	}
+	else if (overlapLog == -1) {
+		overlapLog = 0;
+	}
+
+	TRY_SET_PARAMETER(self->params, ZSTD_c_overlapLog, overlapLog);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_forceMaxWindow, forceMaxWindow);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_enableLongDistanceMatching, enableLDM);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_ldmHashLog, ldmHashLog);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_ldmMinMatch, ldmMinMatch);
+	TRY_SET_PARAMETER(self->params, ZSTD_c_ldmBucketSizeLog, ldmBucketSizeLog);
+
+	if (ldmHashRateLog != -1 && ldmHashEveryLog != -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot specify both ldm_hash_rate_log and ldm_hash_everyLog");
+		return -1;
+	}
+
+	if (ldmHashEveryLog != -1) {
+		ldmHashRateLog = ldmHashEveryLog;
+	}
+	else if (ldmHashRateLog == -1) {
+		ldmHashRateLog = 0;
+	}
+
+	TRY_SET_PARAMETER(self->params, ZSTD_c_ldmHashRateLog, ldmHashRateLog);
+
+	return 0;
+}
+
+PyDoc_STRVAR(ZstdCompressionParameters_from_level__doc__,
+"Create a CompressionParameters from a compression level and target sizes\n"
+);
+
+ZstdCompressionParametersObject* CompressionParameters_from_level(PyObject* undef, PyObject* args, PyObject* kwargs) {
+	int managedKwargs = 0;
+	int level;
+	PyObject* sourceSize = NULL;
+	PyObject* dictSize = NULL;
+	unsigned PY_LONG_LONG iSourceSize = 0;
+	Py_ssize_t iDictSize = 0;
+	PyObject* val;
+	ZSTD_compressionParameters params;
+	ZstdCompressionParametersObject* result = NULL;
+	int res;
+
+	if (!PyArg_ParseTuple(args, "i:from_level",
+		&level)) {
+		return NULL;
+	}
+
+	if (!kwargs) {
+		kwargs = PyDict_New();
+		if (!kwargs) {
+			return NULL;
+		}
+		managedKwargs = 1;
+	}
+
+	sourceSize = PyDict_GetItemString(kwargs, "source_size");
+	if (sourceSize) {
+#if PY_MAJOR_VERSION >= 3
+		iSourceSize = PyLong_AsUnsignedLongLong(sourceSize);
+		if (iSourceSize == (unsigned PY_LONG_LONG)(-1)) {
+			goto cleanup;
+		}
+#else
+		iSourceSize = PyInt_AsUnsignedLongLongMask(sourceSize);
+#endif
+
+		PyDict_DelItemString(kwargs, "source_size");
+	}
+
+	dictSize = PyDict_GetItemString(kwargs, "dict_size");
+	if (dictSize) {
+#if PY_MAJOR_VERSION >= 3
+		iDictSize = PyLong_AsSsize_t(dictSize);
+#else
+		iDictSize = PyInt_AsSsize_t(dictSize);
+#endif
+		if (iDictSize == -1) {
+			goto cleanup;
+		}
+
+		PyDict_DelItemString(kwargs, "dict_size");
+	}
+
+
+	params = ZSTD_getCParams(level, iSourceSize, iDictSize);
+
+	/* Values derived from the input level and sizes are passed along to the
+	   constructor. But only if a value doesn't already exist. */
+	val = PyDict_GetItemString(kwargs, "window_log");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.windowLog);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "window_log", val);
+		Py_DECREF(val);
+	}
+
+	val = PyDict_GetItemString(kwargs, "chain_log");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.chainLog);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "chain_log", val);
+		Py_DECREF(val);
+	}
+
+	val = PyDict_GetItemString(kwargs, "hash_log");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.hashLog);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "hash_log", val);
+		Py_DECREF(val);
+	}
+
+	val = PyDict_GetItemString(kwargs, "search_log");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.searchLog);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "search_log", val);
+		Py_DECREF(val);
+	}
+
+	val = PyDict_GetItemString(kwargs, "min_match");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.minMatch);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "min_match", val);
+		Py_DECREF(val);
+	}
+
+	val = PyDict_GetItemString(kwargs, "target_length");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.targetLength);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "target_length", val);
+		Py_DECREF(val);
+	}
+
+	val = PyDict_GetItemString(kwargs, "compression_strategy");
+	if (!val) {
+		val = PyLong_FromUnsignedLong(params.strategy);
+		if (!val) {
+			goto cleanup;
+		}
+		PyDict_SetItemString(kwargs, "compression_strategy", val);
+		Py_DECREF(val);
+	}
+
+	result = PyObject_New(ZstdCompressionParametersObject, &ZstdCompressionParametersType);
+	if (!result) {
+		goto cleanup;
+	}
+
+	result->params = NULL;
+
+	val = PyTuple_New(0);
+	if (!val) {
+		Py_CLEAR(result);
+		goto cleanup;
+	}
+
+	res = ZstdCompressionParameters_init(result, val, kwargs);
+	Py_DECREF(val);
+
+	if (res) {
+		Py_CLEAR(result);
+		goto cleanup;
+	}
+
+cleanup:
+	if (managedKwargs) {
+		Py_DECREF(kwargs);
+	}
+
+	return result;
+}
+
+PyDoc_STRVAR(ZstdCompressionParameters_estimated_compression_context_size__doc__,
+"Estimate the size in bytes of a compression context for compression parameters\n"
+);
+
+PyObject* ZstdCompressionParameters_estimated_compression_context_size(ZstdCompressionParametersObject* self) {
+	return PyLong_FromSize_t(ZSTD_estimateCCtxSize_usingCCtxParams(self->params));
+}
+
+PyDoc_STRVAR(ZstdCompressionParameters__doc__,
+"ZstdCompressionParameters: low-level control over zstd compression");
+
+static void ZstdCompressionParameters_dealloc(ZstdCompressionParametersObject* self) {
+	if (self->params) {
+		ZSTD_freeCCtxParams(self->params);
+		self->params = NULL;
+	}
+
+	PyObject_Del(self);
+}
+
+#define PARAM_GETTER(name, param) PyObject* ZstdCompressionParameters_get_##name(PyObject* self, void* unused) { \
+    int result; \
+    size_t zresult; \
+    ZstdCompressionParametersObject* p = (ZstdCompressionParametersObject*)(self); \
+    zresult = ZSTD_CCtxParams_getParameter(p->params, param, &result); \
+    if (ZSTD_isError(zresult)) { \
+        PyErr_Format(ZstdError, "unable to get compression parameter: %s", \
+            ZSTD_getErrorName(zresult)); \
+        return NULL; \
+    } \
+    return PyLong_FromLong(result); \
+}
+
+PARAM_GETTER(format, ZSTD_c_format)
+PARAM_GETTER(compression_level, ZSTD_c_compressionLevel)
+PARAM_GETTER(window_log, ZSTD_c_windowLog)
+PARAM_GETTER(hash_log, ZSTD_c_hashLog)
+PARAM_GETTER(chain_log, ZSTD_c_chainLog)
+PARAM_GETTER(search_log, ZSTD_c_searchLog)
+PARAM_GETTER(min_match, ZSTD_c_minMatch)
+PARAM_GETTER(target_length, ZSTD_c_targetLength)
+PARAM_GETTER(compression_strategy, ZSTD_c_strategy)
+PARAM_GETTER(write_content_size, ZSTD_c_contentSizeFlag)
+PARAM_GETTER(write_checksum, ZSTD_c_checksumFlag)
+PARAM_GETTER(write_dict_id, ZSTD_c_dictIDFlag)
+PARAM_GETTER(job_size, ZSTD_c_jobSize)
+PARAM_GETTER(overlap_log, ZSTD_c_overlapLog)
+PARAM_GETTER(force_max_window, ZSTD_c_forceMaxWindow)
+PARAM_GETTER(enable_ldm, ZSTD_c_enableLongDistanceMatching)
+PARAM_GETTER(ldm_hash_log, ZSTD_c_ldmHashLog)
+PARAM_GETTER(ldm_min_match, ZSTD_c_ldmMinMatch)
+PARAM_GETTER(ldm_bucket_size_log, ZSTD_c_ldmBucketSizeLog)
+PARAM_GETTER(ldm_hash_rate_log, ZSTD_c_ldmHashRateLog)
+PARAM_GETTER(threads, ZSTD_c_nbWorkers)
+
+static PyMethodDef ZstdCompressionParameters_methods[] = {
+	{
+		"from_level",
+		(PyCFunction)CompressionParameters_from_level,
+		METH_VARARGS | METH_KEYWORDS | METH_STATIC,
+		ZstdCompressionParameters_from_level__doc__
+	},
+	{
+		"estimated_compression_context_size",
+		(PyCFunction)ZstdCompressionParameters_estimated_compression_context_size,
+		METH_NOARGS,
+		ZstdCompressionParameters_estimated_compression_context_size__doc__
+	},
+	{ NULL, NULL }
+};
+
+#define GET_SET_ENTRY(name) { #name, ZstdCompressionParameters_get_##name, NULL, NULL, NULL }
+
+static PyGetSetDef ZstdCompressionParameters_getset[] = {
+	GET_SET_ENTRY(format),
+	GET_SET_ENTRY(compression_level),
+	GET_SET_ENTRY(window_log),
+	GET_SET_ENTRY(hash_log),
+	GET_SET_ENTRY(chain_log),
+	GET_SET_ENTRY(search_log),
+	GET_SET_ENTRY(min_match),
+	GET_SET_ENTRY(target_length),
+	GET_SET_ENTRY(compression_strategy),
+	GET_SET_ENTRY(write_content_size),
+	GET_SET_ENTRY(write_checksum),
+	GET_SET_ENTRY(write_dict_id),
+	GET_SET_ENTRY(threads),
+	GET_SET_ENTRY(job_size),
+	GET_SET_ENTRY(overlap_log),
+	/* TODO remove this deprecated attribute */
+	{ "overlap_size_log", ZstdCompressionParameters_get_overlap_log, NULL, NULL, NULL },
+	GET_SET_ENTRY(force_max_window),
+	GET_SET_ENTRY(enable_ldm),
+	GET_SET_ENTRY(ldm_hash_log),
+	GET_SET_ENTRY(ldm_min_match),
+	GET_SET_ENTRY(ldm_bucket_size_log),
+	GET_SET_ENTRY(ldm_hash_rate_log),
+	/* TODO remove this deprecated attribute */
+	{ "ldm_hash_every_log", ZstdCompressionParameters_get_ldm_hash_rate_log, NULL, NULL, NULL },
+	{ NULL }
+};
+
+PyTypeObject ZstdCompressionParametersType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"ZstdCompressionParameters", /* tp_name */
+	sizeof(ZstdCompressionParametersObject), /* tp_basicsize */
+	0,                         /* tp_itemsize */
+	(destructor)ZstdCompressionParameters_dealloc, /* tp_dealloc */
+	0,                         /* tp_print */
+	0,                         /* tp_getattr */
+	0,                         /* tp_setattr */
+	0,                         /* tp_compare */
+	0,                         /* tp_repr */
+	0,                         /* tp_as_number */
+	0,                         /* tp_as_sequence */
+	0,                         /* tp_as_mapping */
+	0,                         /* tp_hash  */
+	0,                         /* tp_call */
+	0,                         /* tp_str */
+	0,                         /* tp_getattro */
+	0,                         /* tp_setattro */
+	0,                         /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressionParameters__doc__, /* tp_doc */
+	0,                         /* tp_traverse */
+	0,                         /* tp_clear */
+	0,                         /* tp_richcompare */
+	0,                         /* tp_weaklistoffset */
+	0,                         /* tp_iter */
+	0,                         /* tp_iternext */
+	ZstdCompressionParameters_methods, /* tp_methods */
+	0,                          /* tp_members */
+	ZstdCompressionParameters_getset,  /* tp_getset */
+	0,                         /* tp_base */
+	0,                         /* tp_dict */
+	0,                         /* tp_descr_get */
+	0,                         /* tp_descr_set */
+	0,                         /* tp_dictoffset */
+	(initproc)ZstdCompressionParameters_init, /* tp_init */
+	0,                         /* tp_alloc */
+	PyType_GenericNew,         /* tp_new */
+};
+
+void compressionparams_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdCompressionParametersType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionParametersType) < 0) {
+		return;
+	}
+
+	Py_INCREF(&ZstdCompressionParametersType);
+	PyModule_AddObject(mod, "ZstdCompressionParameters",
+		(PyObject*)&ZstdCompressionParametersType);
+
+	/* TODO remove deprecated alias. */
+	Py_INCREF(&ZstdCompressionParametersType);
+	PyModule_AddObject(mod, "CompressionParameters",
+		(PyObject*)&ZstdCompressionParametersType);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressionreader.c b/contrib/python/zstandard/py2/c-ext/compressionreader.c
new file mode 100644
index 00000000000..47bd3d77053
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressionreader.c
@@ -0,0 +1,818 @@
+/**
+* Copyright (c) 2017-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+static void set_unsupported_operation(void) {
+	PyObject* iomod;
+	PyObject* exc;
+
+	iomod = PyImport_ImportModule("io");
+	if (NULL == iomod) {
+		return;
+	}
+
+	exc = PyObject_GetAttrString(iomod, "UnsupportedOperation");
+	if (NULL == exc) {
+		Py_DECREF(iomod);
+		return;
+	}
+
+	PyErr_SetNone(exc);
+	Py_DECREF(exc);
+	Py_DECREF(iomod);
+}
+
+static void reader_dealloc(ZstdCompressionReader* self) {
+	Py_XDECREF(self->compressor);
+	Py_XDECREF(self->reader);
+
+	if (self->buffer.buf) {
+		PyBuffer_Release(&self->buffer);
+		memset(&self->buffer, 0, sizeof(self->buffer));
+	}
+
+	PyObject_Del(self);
+}
+
+static ZstdCompressionReader* reader_enter(ZstdCompressionReader* self) {
+	if (self->entered) {
+		PyErr_SetString(PyExc_ValueError, "cannot __enter__ multiple times");
+		return NULL;
+	}
+
+	self->entered = 1;
+
+	Py_INCREF(self);
+	return self;
+}
+
+static PyObject* reader_exit(ZstdCompressionReader* self, PyObject* args) {
+	PyObject* exc_type;
+	PyObject* exc_value;
+	PyObject* exc_tb;
+
+	if (!PyArg_ParseTuple(args, "OOO:__exit__", &exc_type, &exc_value, &exc_tb)) {
+		return NULL;
+	}
+
+	self->entered = 0;
+	self->closed = 1;
+
+	/* Release resources associated with source. */
+	Py_CLEAR(self->reader);
+	if (self->buffer.buf) {
+		PyBuffer_Release(&self->buffer);
+		memset(&self->buffer, 0, sizeof(self->buffer));
+	}
+
+    Py_CLEAR(self->compressor);
+
+	Py_RETURN_FALSE;
+}
+
+static PyObject* reader_readable(ZstdCompressionReader* self) {
+	Py_RETURN_TRUE;
+}
+
+static PyObject* reader_writable(ZstdCompressionReader* self) {
+	Py_RETURN_FALSE;
+}
+
+static PyObject* reader_seekable(ZstdCompressionReader* self) {
+	Py_RETURN_FALSE;
+}
+
+static PyObject* reader_readline(PyObject* self, PyObject* args) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_readlines(PyObject* self, PyObject* args) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_write(PyObject* self, PyObject* args) {
+	PyErr_SetString(PyExc_OSError, "stream is not writable");
+	return NULL;
+}
+
+static PyObject* reader_writelines(PyObject* self, PyObject* args) {
+	PyErr_SetString(PyExc_OSError, "stream is not writable");
+	return NULL;
+}
+
+static PyObject* reader_isatty(PyObject* self) {
+	Py_RETURN_FALSE;
+}
+
+static PyObject* reader_flush(PyObject* self) {
+	Py_RETURN_NONE;
+}
+
+static PyObject* reader_close(ZstdCompressionReader* self) {
+	self->closed = 1;
+	Py_RETURN_NONE;
+}
+
+static PyObject* reader_tell(ZstdCompressionReader* self) {
+	/* TODO should this raise OSError since stream isn't seekable? */
+	return PyLong_FromUnsignedLongLong(self->bytesCompressed);
+}
+
+int read_compressor_input(ZstdCompressionReader* self) {
+	if (self->finishedInput) {
+		return 0;
+	}
+
+	if (self->input.pos != self->input.size) {
+		return 0;
+	}
+
+	if (self->reader) {
+		Py_buffer buffer;
+
+		assert(self->readResult == NULL);
+
+		self->readResult = PyObject_CallMethod(self->reader, "read",
+		    "k", self->readSize);
+
+		if (NULL == self->readResult) {
+			return -1;
+		}
+
+		memset(&buffer, 0, sizeof(buffer));
+
+		if (0 != PyObject_GetBuffer(self->readResult, &buffer, PyBUF_CONTIG_RO)) {
+			return -1;
+		}
+
+		/* EOF */
+		if (0 == buffer.len) {
+			self->finishedInput = 1;
+			Py_CLEAR(self->readResult);
+		}
+		else {
+			self->input.src = buffer.buf;
+			self->input.size = buffer.len;
+			self->input.pos = 0;
+		}
+
+		PyBuffer_Release(&buffer);
+	}
+	else {
+		assert(self->buffer.buf);
+
+		self->input.src = self->buffer.buf;
+		self->input.size = self->buffer.len;
+		self->input.pos = 0;
+	}
+
+	return 1;
+}
+
+int compress_input(ZstdCompressionReader* self, ZSTD_outBuffer* output) {
+	size_t oldPos;
+	size_t zresult;
+
+	/* If we have data left over, consume it. */
+	if (self->input.pos < self->input.size) {
+		oldPos = output->pos;
+
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(self->compressor->cctx,
+		    output, &self->input, ZSTD_e_continue);
+		Py_END_ALLOW_THREADS
+
+		self->bytesCompressed += output->pos - oldPos;
+
+		/* Input exhausted. Clear out state tracking. */
+		if (self->input.pos == self->input.size) {
+			memset(&self->input, 0, sizeof(self->input));
+			Py_CLEAR(self->readResult);
+
+			if (self->buffer.buf) {
+				self->finishedInput = 1;
+			}
+		}
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+			return -1;
+		}
+	}
+
+    if (output->pos && output->pos == output->size) {
+        return 1;
+    }
+    else {
+        return 0;
+    }
+}
+
+static PyObject* reader_read(ZstdCompressionReader* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"size",
+		NULL
+	};
+
+	Py_ssize_t size = -1;
+	PyObject* result = NULL;
+	char* resultBuffer;
+	Py_ssize_t resultSize;
+	size_t zresult;
+	size_t oldPos;
+	int readResult, compressResult;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n", kwlist, &size)) {
+		return NULL;
+	}
+
+	if (size < -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot read negative amounts less than -1");
+		return NULL;
+	}
+
+	if (size == -1) {
+		return PyObject_CallMethod((PyObject*)self, "readall", NULL);
+	}
+
+	if (self->finishedOutput || size == 0) {
+		return PyBytes_FromStringAndSize("", 0);
+	}
+
+	result = PyBytes_FromStringAndSize(NULL, size);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+	self->output.dst = resultBuffer;
+	self->output.size = resultSize;
+	self->output.pos = 0;
+
+readinput:
+
+    compressResult = compress_input(self, &self->output);
+
+	if (-1 == compressResult) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+	else if (0 == compressResult) {
+		/* There is room in the output. We fall through to below, which will
+		 * either get more input for us or will attempt to end the stream.
+		 */
+	}
+	else if (1 == compressResult) {
+		memset(&self->output, 0, sizeof(self->output));
+		return result;
+	}
+	else {
+		assert(0);
+	}
+
+	readResult = read_compressor_input(self);
+
+	if (-1 == readResult) {
+		return NULL;
+	}
+	else if (0 == readResult) { }
+	else if (1 == readResult) { }
+	else {
+		assert(0);
+	}
+
+	if (self->input.size) {
+		goto readinput;
+	}
+
+	/* Else EOF */
+	oldPos = self->output.pos;
+
+	zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+		&self->input, ZSTD_e_end);
+
+	self->bytesCompressed += self->output.pos - oldPos;
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error ending compression stream: %s",
+			ZSTD_getErrorName(zresult));
+		Py_XDECREF(result);
+		return NULL;
+	}
+
+	assert(self->output.pos);
+
+	if (0 == zresult) {
+		self->finishedOutput = 1;
+	}
+
+	if (safe_pybytes_resize(&result, self->output.pos)) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+
+	memset(&self->output, 0, sizeof(self->output));
+
+	return result;
+}
+
+static PyObject* reader_read1(ZstdCompressionReader* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"size",
+		NULL
+	};
+
+	Py_ssize_t size = -1;
+	PyObject* result = NULL;
+	char* resultBuffer;
+	Py_ssize_t resultSize;
+	ZSTD_outBuffer output;
+	int compressResult;
+	size_t oldPos;
+	size_t zresult;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n:read1", kwlist, &size)) {
+		return NULL;
+	}
+
+	if (size < -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot read negative amounts less than -1");
+		return NULL;
+	}
+
+	if (self->finishedOutput || size == 0) {
+		return PyBytes_FromStringAndSize("", 0);
+	}
+
+	if (size == -1) {
+		size = ZSTD_CStreamOutSize();
+	}
+
+	result = PyBytes_FromStringAndSize(NULL, size);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+	output.dst = resultBuffer;
+	output.size = resultSize;
+	output.pos = 0;
+
+	/* read1() is supposed to use at most 1 read() from the underlying stream.
+	   However, we can't satisfy this requirement with compression because
+	   not every input will generate output. We /could/ flush the compressor,
+	   but this may not be desirable. We allow multiple read() from the
+	   underlying stream. But unlike read(), we return as soon as output data
+	   is available.
+	*/
+
+	compressResult = compress_input(self, &output);
+
+	if (-1 == compressResult) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+	else if (0 == compressResult || 1 == compressResult) { }
+	else {
+		assert(0);
+	}
+
+	if (output.pos) {
+		goto finally;
+	}
+
+	while (!self->finishedInput) {
+		int readResult = read_compressor_input(self);
+
+		if (-1 == readResult) {
+			Py_XDECREF(result);
+			return NULL;
+		}
+		else if (0 == readResult || 1 == readResult) { }
+		else {
+			assert(0);
+		}
+
+		compressResult = compress_input(self, &output);
+
+		if (-1 == compressResult) {
+			Py_XDECREF(result);
+			return NULL;
+		}
+		else if (0 == compressResult || 1 == compressResult) { }
+		else {
+			assert(0);
+		}
+
+		if (output.pos) {
+			goto finally;
+		}
+	}
+
+	/* EOF */
+	oldPos = output.pos;
+
+	zresult = ZSTD_compressStream2(self->compressor->cctx, &output, &self->input,
+        ZSTD_e_end);
+
+	self->bytesCompressed += output.pos - oldPos;
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error ending compression stream: %s",
+		    ZSTD_getErrorName(zresult));
+		Py_XDECREF(result);
+		return NULL;
+	}
+
+	if (zresult == 0) {
+		self->finishedOutput = 1;
+	}
+
+finally:
+	if (result) {
+		if (safe_pybytes_resize(&result, output.pos)) {
+			Py_XDECREF(result);
+			return NULL;
+		}
+	}
+
+	return result;
+}
+
+static PyObject* reader_readall(PyObject* self) {
+	PyObject* chunks = NULL;
+	PyObject* empty = NULL;
+	PyObject* result = NULL;
+
+	/* Our strategy is to collect chunks into a list then join all the
+	 * chunks at the end. We could potentially use e.g. an io.BytesIO. But
+	 * this feels simple enough to implement and avoids potentially expensive
+	 * reallocations of large buffers.
+	 */
+	chunks = PyList_New(0);
+	if (NULL == chunks) {
+		return NULL;
+	}
+
+	while (1) {
+		PyObject* chunk = PyObject_CallMethod(self, "read", "i", 1048576);
+		if (NULL == chunk) {
+			Py_DECREF(chunks);
+			return NULL;
+		}
+
+		if (!PyBytes_Size(chunk)) {
+			Py_DECREF(chunk);
+			break;
+		}
+
+		if (PyList_Append(chunks, chunk)) {
+			Py_DECREF(chunk);
+			Py_DECREF(chunks);
+			return NULL;
+		}
+
+		Py_DECREF(chunk);
+	}
+
+	empty = PyBytes_FromStringAndSize("", 0);
+	if (NULL == empty) {
+		Py_DECREF(chunks);
+		return NULL;
+	}
+
+	result = PyObject_CallMethod(empty, "join", "O", chunks);
+
+	Py_DECREF(empty);
+	Py_DECREF(chunks);
+
+	return result;
+}
+
+static PyObject* reader_readinto(ZstdCompressionReader* self, PyObject* args) {
+	Py_buffer dest;
+	ZSTD_outBuffer output;
+	int readResult, compressResult;
+	PyObject* result = NULL;
+	size_t zresult;
+	size_t oldPos;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (self->finishedOutput) {
+		return PyLong_FromLong(0);
+	}
+
+	if (!PyArg_ParseTuple(args, "w*:readinto", &dest)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&dest, 'C') || dest.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+		    "destination buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	output.dst = dest.buf;
+	output.size = dest.len;
+	output.pos = 0;
+
+	compressResult = compress_input(self, &output);
+
+	if (-1 == compressResult) {
+		goto finally;
+	}
+	else if (0 == compressResult) {	}
+	else if (1 == compressResult) {
+		result = PyLong_FromSize_t(output.pos);
+		goto finally;
+	}
+	else {
+		assert(0);
+	}
+
+	while (!self->finishedInput) {
+		readResult = read_compressor_input(self);
+
+		if (-1 == readResult) {
+			goto finally;
+		}
+		else if (0 == readResult || 1 == readResult) {}
+		else {
+			assert(0);
+		}
+
+		compressResult = compress_input(self, &output);
+
+		if (-1 == compressResult) {
+			goto finally;
+		}
+		else if (0 == compressResult) { }
+		else if (1 == compressResult) {
+			result = PyLong_FromSize_t(output.pos);
+			goto finally;
+		}
+		else {
+			assert(0);
+		}
+	}
+
+	/* EOF */
+	oldPos = output.pos;
+
+	zresult = ZSTD_compressStream2(self->compressor->cctx, &output, &self->input,
+	    ZSTD_e_end);
+
+	self->bytesCompressed += self->output.pos - oldPos;
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error ending compression stream: %s",
+		    ZSTD_getErrorName(zresult));
+		goto finally;
+	}
+
+	assert(output.pos);
+
+	if (0 == zresult) {
+		self->finishedOutput = 1;
+	}
+
+	result = PyLong_FromSize_t(output.pos);
+
+finally:
+	PyBuffer_Release(&dest);
+
+	return result;
+}
+
+static PyObject* reader_readinto1(ZstdCompressionReader* self, PyObject* args) {
+	Py_buffer dest;
+	PyObject* result = NULL;
+	ZSTD_outBuffer output;
+	int compressResult;
+	size_t oldPos;
+	size_t zresult;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (self->finishedOutput) {
+		return PyLong_FromLong(0);
+	}
+
+	if (!PyArg_ParseTuple(args, "w*:readinto1", &dest)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&dest, 'C') || dest.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+		    "destination buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	output.dst = dest.buf;
+	output.size = dest.len;
+	output.pos = 0;
+
+	compressResult = compress_input(self, &output);
+
+	if (-1 == compressResult) {
+		goto finally;
+	}
+	else if (0 == compressResult || 1 == compressResult) { }
+	else {
+		assert(0);
+	}
+
+	if (output.pos) {
+		result = PyLong_FromSize_t(output.pos);
+		goto finally;
+	}
+
+	while (!self->finishedInput) {
+		int readResult = read_compressor_input(self);
+
+		if (-1 == readResult) {
+			goto finally;
+		}
+		else if (0 == readResult || 1 == readResult) { }
+		else {
+			assert(0);
+		}
+
+		compressResult = compress_input(self, &output);
+
+		if (-1 == compressResult) {
+			goto finally;
+		}
+		else if (0 == compressResult) { }
+		else if (1 == compressResult) {
+			result = PyLong_FromSize_t(output.pos);
+			goto finally;
+		}
+		else {
+			assert(0);
+		}
+
+		/* If we produced output and we're not done with input, emit
+		 * that output now, as we've hit restrictions of read1().
+		 */
+		if (output.pos && !self->finishedInput) {
+			result = PyLong_FromSize_t(output.pos);
+			goto finally;
+		}
+
+		/* Otherwise we either have no output or we've exhausted the
+		 * input. Either we try to get more input or we fall through
+		 * to EOF below */
+	}
+
+	/* EOF */
+	oldPos = output.pos;
+
+	zresult = ZSTD_compressStream2(self->compressor->cctx, &output, &self->input,
+	    ZSTD_e_end);
+
+	self->bytesCompressed += self->output.pos - oldPos;
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error ending compression stream: %s",
+		    ZSTD_getErrorName(zresult));
+		goto finally;
+	}
+
+	assert(output.pos);
+
+	if (0 == zresult) {
+		self->finishedOutput = 1;
+	}
+
+	result = PyLong_FromSize_t(output.pos);
+
+finally:
+	PyBuffer_Release(&dest);
+
+	return result;
+}
+
+static PyObject* reader_iter(PyObject* self) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_iternext(PyObject* self) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyMethodDef reader_methods[] = {
+	{ "__enter__", (PyCFunction)reader_enter, METH_NOARGS,
+	PyDoc_STR("Enter a compression context") },
+	{ "__exit__", (PyCFunction)reader_exit, METH_VARARGS,
+	PyDoc_STR("Exit a compression context") },
+	{ "close", (PyCFunction)reader_close, METH_NOARGS,
+	PyDoc_STR("Close the stream so it cannot perform any more operations") },
+	{ "flush", (PyCFunction)reader_flush, METH_NOARGS, PyDoc_STR("no-ops") },
+	{ "isatty", (PyCFunction)reader_isatty, METH_NOARGS, PyDoc_STR("Returns False") },
+	{ "readable", (PyCFunction)reader_readable, METH_NOARGS,
+	PyDoc_STR("Returns True") },
+	{ "read", (PyCFunction)reader_read, METH_VARARGS | METH_KEYWORDS, PyDoc_STR("read compressed data") },
+	{ "read1", (PyCFunction)reader_read1, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readall", (PyCFunction)reader_readall, METH_NOARGS, PyDoc_STR("Not implemented") },
+	{ "readinto", (PyCFunction)reader_readinto, METH_VARARGS, NULL },
+	{ "readinto1", (PyCFunction)reader_readinto1, METH_VARARGS, NULL },
+	{ "readline", (PyCFunction)reader_readline, METH_VARARGS, PyDoc_STR("Not implemented") },
+	{ "readlines", (PyCFunction)reader_readlines, METH_VARARGS, PyDoc_STR("Not implemented") },
+	{ "seekable", (PyCFunction)reader_seekable, METH_NOARGS,
+	PyDoc_STR("Returns False") },
+	{ "tell", (PyCFunction)reader_tell, METH_NOARGS,
+	PyDoc_STR("Returns current number of bytes compressed") },
+	{ "writable", (PyCFunction)reader_writable, METH_NOARGS,
+	PyDoc_STR("Returns False") },
+	{ "write", reader_write, METH_VARARGS, PyDoc_STR("Raises OSError") },
+	{ "writelines", reader_writelines, METH_VARARGS, PyDoc_STR("Not implemented") },
+	{ NULL, NULL }
+};
+
+static PyMemberDef reader_members[] = {
+	{ "closed", T_BOOL, offsetof(ZstdCompressionReader, closed),
+	  READONLY, "whether stream is closed" },
+	{ NULL }
+};
+
+PyTypeObject ZstdCompressionReaderType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressionReader", /* tp_name */
+	sizeof(ZstdCompressionReader), /* tp_basicsize */
+	0, /* tp_itemsize */
+	(destructor)reader_dealloc, /* tp_dealloc */
+	0, /* tp_print */
+	0, /* tp_getattr */
+	0, /* tp_setattr */
+	0, /* tp_compare */
+	0, /* tp_repr */
+	0, /* tp_as_number */
+	0, /* tp_as_sequence */
+	0, /* tp_as_mapping */
+	0, /* tp_hash */
+	0, /* tp_call */
+	0, /* tp_str */
+	0, /* tp_getattro */
+	0, /* tp_setattro */
+	0, /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT, /* tp_flags */
+	0, /* tp_doc */
+	0, /* tp_traverse */
+	0, /* tp_clear */
+	0, /* tp_richcompare */
+	0, /* tp_weaklistoffset */
+	reader_iter, /* tp_iter */
+	reader_iternext, /* tp_iternext */
+	reader_methods, /* tp_methods */
+	reader_members, /* tp_members */
+	0, /* tp_getset */
+	0, /* tp_base */
+	0, /* tp_dict */
+	0, /* tp_descr_get */
+	0, /* tp_descr_set */
+	0, /* tp_dictoffset */
+	0, /* tp_init */
+	0, /* tp_alloc */
+	PyType_GenericNew, /* tp_new */
+};
+
+void compressionreader_module_init(PyObject* mod) {
+	/* TODO make reader a sub-class of io.RawIOBase */
+
+	Py_TYPE(&ZstdCompressionReaderType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionReaderType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressionwriter.c b/contrib/python/zstandard/py2/c-ext/compressionwriter.c
new file mode 100644
index 00000000000..fe8e55a0fe0
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressionwriter.c
@@ -0,0 +1,372 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(ZstdCompresssionWriter__doc__,
+"""A context manager used for writing compressed output to a writer.\n"
+);
+
+static void ZstdCompressionWriter_dealloc(ZstdCompressionWriter* self) {
+	Py_XDECREF(self->compressor);
+	Py_XDECREF(self->writer);
+
+	PyMem_Free(self->output.dst);
+	self->output.dst = NULL;
+
+	PyObject_Del(self);
+}
+
+static PyObject* ZstdCompressionWriter_enter(ZstdCompressionWriter* self) {
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (self->entered) {
+		PyErr_SetString(ZstdError, "cannot __enter__ multiple times");
+		return NULL;
+	}
+
+	self->entered = 1;
+
+	Py_INCREF(self);
+	return (PyObject*)self;
+}
+
+static PyObject* ZstdCompressionWriter_exit(ZstdCompressionWriter* self, PyObject* args) {
+	PyObject* exc_type;
+	PyObject* exc_value;
+	PyObject* exc_tb;
+
+	if (!PyArg_ParseTuple(args, "OOO:__exit__", &exc_type, &exc_value, &exc_tb)) {
+		return NULL;
+	}
+
+	self->entered = 0;
+
+	if (exc_type == Py_None && exc_value == Py_None && exc_tb == Py_None) {
+		PyObject* result = PyObject_CallMethod((PyObject*)self, "close", NULL);
+
+		if (NULL == result) {
+			return NULL;
+		}
+	}
+
+	Py_RETURN_FALSE;
+}
+
+static PyObject* ZstdCompressionWriter_memory_size(ZstdCompressionWriter* self) {
+	return PyLong_FromSize_t(ZSTD_sizeof_CCtx(self->compressor->cctx));
+}
+
+static PyObject* ZstdCompressionWriter_write(ZstdCompressionWriter* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	PyObject* result = NULL;
+	Py_buffer source;
+	size_t zresult;
+	ZSTD_inBuffer input;
+	PyObject* res;
+	Py_ssize_t totalWrite = 0;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:write",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:write",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	self->output.pos = 0;
+
+	input.src = source.buf;
+	input.size = source.len;
+	input.pos = 0;
+
+	while (input.pos < (size_t)source.len) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output, &input, ZSTD_e_continue);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+			goto finally;
+		}
+
+		/* Copy data from output buffer to writer. */
+		if (self->output.pos) {
+#if PY_MAJOR_VERSION >= 3
+			res = PyObject_CallMethod(self->writer, "write", "y#",
+#else
+			res = PyObject_CallMethod(self->writer, "write", "s#",
+#endif
+				self->output.dst, self->output.pos);
+			Py_XDECREF(res);
+			totalWrite += self->output.pos;
+			self->bytesCompressed += self->output.pos;
+		}
+		self->output.pos = 0;
+	}
+
+	if (self->writeReturnRead) {
+		result = PyLong_FromSize_t(input.pos);
+	}
+	else {
+		result = PyLong_FromSsize_t(totalWrite);
+	}
+
+finally:
+	PyBuffer_Release(&source);
+	return result;
+}
+
+static PyObject* ZstdCompressionWriter_flush(ZstdCompressionWriter* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"flush_mode",
+		NULL
+	};
+
+	size_t zresult;
+	ZSTD_inBuffer input;
+	PyObject* res;
+	Py_ssize_t totalWrite = 0;
+	unsigned flush_mode = 0;
+	ZSTD_EndDirective flush;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|I:flush",
+		kwlist, &flush_mode)) {
+		return NULL;
+	}
+
+	switch (flush_mode) {
+		case 0:
+			flush = ZSTD_e_flush;
+			break;
+		case 1:
+			flush = ZSTD_e_end;
+			break;
+		default:
+			PyErr_Format(PyExc_ValueError, "unknown flush_mode: %d", flush_mode);
+			return NULL;
+	}
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	self->output.pos = 0;
+
+	input.src = NULL;
+	input.size = 0;
+	input.pos = 0;
+
+	while (1) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output, &input, flush);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+			return NULL;
+		}
+
+		/* Copy data from output buffer to writer. */
+		if (self->output.pos) {
+#if PY_MAJOR_VERSION >= 3
+			res = PyObject_CallMethod(self->writer, "write", "y#",
+#else
+			res = PyObject_CallMethod(self->writer, "write", "s#",
+#endif
+				self->output.dst, self->output.pos);
+			Py_XDECREF(res);
+			totalWrite += self->output.pos;
+			self->bytesCompressed += self->output.pos;
+		}
+
+		self->output.pos = 0;
+
+		if (!zresult) {
+			break;
+		}
+	}
+
+	return PyLong_FromSsize_t(totalWrite);
+}
+
+static PyObject* ZstdCompressionWriter_close(ZstdCompressionWriter* self) {
+	PyObject* result;
+
+	if (self->closed) {
+		Py_RETURN_NONE;
+	}
+
+	result = PyObject_CallMethod((PyObject*)self, "flush", "I", 1);
+	self->closed = 1;
+
+	if (NULL == result) {
+	    return NULL;
+	}
+
+    /* Call close on underlying stream as well. */
+	if (PyObject_HasAttrString(self->writer, "close")) {
+		return PyObject_CallMethod(self->writer, "close", NULL);
+	}
+
+	Py_RETURN_NONE;
+}
+
+static PyObject* ZstdCompressionWriter_fileno(ZstdCompressionWriter* self) {
+	if (PyObject_HasAttrString(self->writer, "fileno")) {
+		return PyObject_CallMethod(self->writer, "fileno", NULL);
+	}
+	else {
+		PyErr_SetString(PyExc_OSError, "fileno not available on underlying writer");
+		return NULL;
+	}
+}
+
+static PyObject* ZstdCompressionWriter_tell(ZstdCompressionWriter* self) {
+	return PyLong_FromUnsignedLongLong(self->bytesCompressed);
+}
+
+static PyObject* ZstdCompressionWriter_writelines(PyObject* self, PyObject* args) {
+	PyErr_SetNone(PyExc_NotImplementedError);
+	return NULL;
+}
+
+static PyObject* ZstdCompressionWriter_false(PyObject* self, PyObject* args) {
+	Py_RETURN_FALSE;
+}
+
+static PyObject* ZstdCompressionWriter_true(PyObject* self, PyObject* args) {
+	Py_RETURN_TRUE;
+}
+
+static PyObject* ZstdCompressionWriter_unsupported(PyObject* self, PyObject* args, PyObject* kwargs) {
+	PyObject* iomod;
+	PyObject* exc;
+
+	iomod = PyImport_ImportModule("io");
+	if (NULL == iomod) {
+		return NULL;
+	}
+
+	exc = PyObject_GetAttrString(iomod, "UnsupportedOperation");
+	if (NULL == exc) {
+		Py_DECREF(iomod);
+		return NULL;
+	}
+
+	PyErr_SetNone(exc);
+	Py_DECREF(exc);
+	Py_DECREF(iomod);
+
+	return NULL;
+}
+
+static PyMethodDef ZstdCompressionWriter_methods[] = {
+	{ "__enter__", (PyCFunction)ZstdCompressionWriter_enter, METH_NOARGS,
+	PyDoc_STR("Enter a compression context.") },
+	{ "__exit__", (PyCFunction)ZstdCompressionWriter_exit, METH_VARARGS,
+	PyDoc_STR("Exit a compression context.") },
+	{ "close", (PyCFunction)ZstdCompressionWriter_close, METH_NOARGS, NULL },
+	{ "fileno", (PyCFunction)ZstdCompressionWriter_fileno, METH_NOARGS, NULL },
+	{ "isatty", (PyCFunction)ZstdCompressionWriter_false, METH_NOARGS, NULL },
+	{ "readable", (PyCFunction)ZstdCompressionWriter_false, METH_NOARGS, NULL },
+	{ "readline", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readlines", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "seek", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "seekable", ZstdCompressionWriter_false, METH_NOARGS, NULL },
+	{ "truncate", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "writable", ZstdCompressionWriter_true, METH_NOARGS, NULL },
+	{ "writelines", ZstdCompressionWriter_writelines, METH_VARARGS, NULL },
+	{ "read", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readall", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readinto", (PyCFunction)ZstdCompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "memory_size", (PyCFunction)ZstdCompressionWriter_memory_size, METH_NOARGS,
+	PyDoc_STR("Obtain the memory size of the underlying compressor") },
+	{ "write", (PyCFunction)ZstdCompressionWriter_write, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("Compress data") },
+	{ "flush", (PyCFunction)ZstdCompressionWriter_flush, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("Flush data and finish a zstd frame") },
+	{ "tell", (PyCFunction)ZstdCompressionWriter_tell, METH_NOARGS,
+	PyDoc_STR("Returns current number of bytes compressed") },
+	{ NULL, NULL }
+};
+
+static PyMemberDef ZstdCompressionWriter_members[] = {
+	 { "closed", T_BOOL, offsetof(ZstdCompressionWriter, closed), READONLY, NULL },
+	 { NULL }
+};
+
+PyTypeObject ZstdCompressionWriterType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressionWriter",  /* tp_name */
+	sizeof(ZstdCompressionWriter),  /* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)ZstdCompressionWriter_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	0,                              /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompresssionWriter__doc__,  /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	ZstdCompressionWriter_methods,  /* tp_methods */
+	ZstdCompressionWriter_members,  /* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	0,                              /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void compressionwriter_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdCompressionWriterType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionWriterType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressobj.c b/contrib/python/zstandard/py2/c-ext/compressobj.c
new file mode 100644
index 00000000000..0c33dffc7af
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressobj.c
@@ -0,0 +1,256 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(ZstdCompressionObj__doc__,
+"Perform compression using a standard library compatible API.\n"
+);
+
+static void ZstdCompressionObj_dealloc(ZstdCompressionObj* self) {
+	PyMem_Free(self->output.dst);
+	self->output.dst = NULL;
+
+	Py_XDECREF(self->compressor);
+
+	PyObject_Del(self);
+}
+
+static PyObject* ZstdCompressionObj_compress(ZstdCompressionObj* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	Py_buffer source;
+	ZSTD_inBuffer input;
+	size_t zresult;
+	PyObject* result = NULL;
+	Py_ssize_t resultSize = 0;
+
+	if (self->finished) {
+		PyErr_SetString(ZstdError, "cannot call compress() after compressor finished");
+		return NULL;
+	}
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:compress",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:compress",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	input.src = source.buf;
+	input.size = source.len;
+	input.pos = 0;
+
+	while (input.pos < (size_t)source.len) {
+		Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+				&input, ZSTD_e_continue);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+			Py_CLEAR(result);
+			goto finally;
+		}
+
+		if (self->output.pos) {
+			if (result) {
+				resultSize = PyBytes_GET_SIZE(result);
+
+				if (safe_pybytes_resize(&result, resultSize + self->output.pos)) {
+					Py_CLEAR(result);
+					goto finally;
+				}
+
+				memcpy(PyBytes_AS_STRING(result) + resultSize,
+					self->output.dst, self->output.pos);
+			}
+			else {
+				result = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+				if (!result) {
+					goto finally;
+				}
+			}
+
+			self->output.pos = 0;
+		}
+	}
+
+	if (NULL == result) {
+		result = PyBytes_FromString("");
+	}
+
+finally:
+	PyBuffer_Release(&source);
+
+	return result;
+}
+
+static PyObject* ZstdCompressionObj_flush(ZstdCompressionObj* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"flush_mode",
+		NULL
+	};
+
+	int flushMode = compressorobj_flush_finish;
+	size_t zresult;
+	PyObject* result = NULL;
+	Py_ssize_t resultSize = 0;
+	ZSTD_inBuffer input;
+	ZSTD_EndDirective zFlushMode;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|i:flush", kwlist, &flushMode)) {
+		return NULL;
+	}
+
+	if (flushMode != compressorobj_flush_finish && flushMode != compressorobj_flush_block) {
+		PyErr_SetString(PyExc_ValueError, "flush mode not recognized");
+		return NULL;
+	}
+
+	if (self->finished) {
+		PyErr_SetString(ZstdError, "compressor object already finished");
+		return NULL;
+	}
+
+	switch (flushMode) {
+		case compressorobj_flush_block:
+			zFlushMode = ZSTD_e_flush;
+			break;
+
+		case compressorobj_flush_finish:
+			zFlushMode = ZSTD_e_end;
+			self->finished = 1;
+			break;
+
+		default:
+			PyErr_SetString(ZstdError, "unhandled flush mode");
+			return NULL;
+	}
+
+	assert(self->output.pos == 0);
+
+	input.src = NULL;
+	input.size = 0;
+	input.pos = 0;
+
+	while (1) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+			&input, zFlushMode);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "error ending compression stream: %s",
+				ZSTD_getErrorName(zresult));
+			return NULL;
+		}
+
+		if (self->output.pos) {
+			if (result) {
+				resultSize = PyBytes_GET_SIZE(result);
+
+				if (safe_pybytes_resize(&result, resultSize + self->output.pos)) {
+					Py_XDECREF(result);
+					return NULL;
+				}
+
+				memcpy(PyBytes_AS_STRING(result) + resultSize,
+					self->output.dst, self->output.pos);
+			}
+			else {
+				result = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+				if (!result) {
+					return NULL;
+				}
+			}
+
+			self->output.pos = 0;
+		}
+
+		if (!zresult) {
+			break;
+		}
+	}
+
+	if (result) {
+		return result;
+	}
+	else {
+		return PyBytes_FromString("");
+	}
+}
+
+static PyMethodDef ZstdCompressionObj_methods[] = {
+	{ "compress", (PyCFunction)ZstdCompressionObj_compress, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("compress data") },
+	{ "flush", (PyCFunction)ZstdCompressionObj_flush, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("finish compression operation") },
+	{ NULL, NULL }
+};
+
+PyTypeObject ZstdCompressionObjType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressionObj",      /* tp_name */
+	sizeof(ZstdCompressionObj),     /* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)ZstdCompressionObj_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	0,                              /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressionObj__doc__,      /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	ZstdCompressionObj_methods,     /* tp_methods */
+	0,                              /* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	0,                              /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void compressobj_module_init(PyObject* module) {
+	Py_TYPE(&ZstdCompressionObjType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressionObjType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressor.c b/contrib/python/zstandard/py2/c-ext/compressor.c
new file mode 100644
index 00000000000..fb729c17c4e
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressor.c
@@ -0,0 +1,1676 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+#include "pool.h"
+
+extern PyObject* ZstdError;
+
+int setup_cctx(ZstdCompressor* compressor) {
+	size_t zresult;
+
+	assert(compressor);
+	assert(compressor->cctx);
+	assert(compressor->params);
+
+	zresult = ZSTD_CCtx_setParametersUsingCCtxParams(compressor->cctx, compressor->params);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "could not set compression parameters: %s",
+			ZSTD_getErrorName(zresult));
+		return 1;
+	}
+
+	if (compressor->dict) {
+		if (compressor->dict->cdict) {
+			zresult = ZSTD_CCtx_refCDict(compressor->cctx, compressor->dict->cdict);
+		}
+		else {
+			zresult = ZSTD_CCtx_loadDictionary_advanced(compressor->cctx,
+				compressor->dict->dictData, compressor->dict->dictSize,
+				ZSTD_dlm_byRef, compressor->dict->dictType);
+		}
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "could not load compression dictionary: %s",
+				ZSTD_getErrorName(zresult));
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+static PyObject* frame_progression(ZSTD_CCtx* cctx) {
+	PyObject* result = NULL;
+	PyObject* value;
+	ZSTD_frameProgression progression;
+
+	result = PyTuple_New(3);
+	if (!result) {
+		return NULL;
+	}
+
+	progression = ZSTD_getFrameProgression(cctx);
+
+	value = PyLong_FromUnsignedLongLong(progression.ingested);
+	if (!value) {
+		Py_DECREF(result);
+		return NULL;
+	}
+
+	PyTuple_SET_ITEM(result, 0, value);
+
+	value = PyLong_FromUnsignedLongLong(progression.consumed);
+	if (!value) {
+		Py_DECREF(result);
+		return NULL;
+	}
+
+	PyTuple_SET_ITEM(result, 1, value);
+
+	value = PyLong_FromUnsignedLongLong(progression.produced);
+	if (!value) {
+		Py_DECREF(result);
+		return NULL;
+	}
+
+	PyTuple_SET_ITEM(result, 2, value);
+
+	return result;
+}
+
+PyDoc_STRVAR(ZstdCompressor__doc__,
+"ZstdCompressor(level=None, dict_data=None, compression_params=None)\n"
+"\n"
+"Create an object used to perform Zstandard compression.\n"
+"\n"
+"An instance can compress data various ways. Instances can be used multiple\n"
+"times. Each compression operation will use the compression parameters\n"
+"defined at construction time.\n"
+"\n"
+"Compression can be configured via the following names arguments:\n"
+"\n"
+"level\n"
+"   Integer compression level.\n"
+"dict_data\n"
+"   A ``ZstdCompressionDict`` to be used to compress with dictionary data.\n"
+"compression_params\n"
+"   A ``CompressionParameters`` instance defining low-level compression"
+"   parameters. If defined, this will overwrite the ``level`` argument.\n"
+"write_checksum\n"
+"   If True, a 4 byte content checksum will be written with the compressed\n"
+"   data, allowing the decompressor to perform content verification.\n"
+"write_content_size\n"
+"   If True (the default), the decompressed content size will be included in\n"
+"   the header of the compressed data. This data will only be written if the\n"
+"   compressor knows the size of the input data.\n"
+"write_dict_id\n"
+"   Determines whether the dictionary ID will be written into the compressed\n"
+"   data. Defaults to True. Only adds content to the compressed data if\n"
+"   a dictionary is being used.\n"
+"threads\n"
+"   Number of threads to use to compress data concurrently. When set,\n"
+"   compression operations are performed on multiple threads. The default\n"
+"   value (0) disables multi-threaded compression. A value of ``-1`` means to\n"
+"   set the number of threads to the number of detected logical CPUs.\n"
+);
+
+static int ZstdCompressor_init(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"level",
+		"dict_data",
+		"compression_params",
+		"write_checksum",
+		"write_content_size",
+		"write_dict_id",
+		"threads",
+		NULL
+	};
+
+	int level = 3;
+	ZstdCompressionDict* dict = NULL;
+	ZstdCompressionParametersObject* params = NULL;
+	PyObject* writeChecksum = NULL;
+	PyObject* writeContentSize = NULL;
+	PyObject* writeDictID = NULL;
+	int threads = 0;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|iO!O!OOOi:ZstdCompressor",
+		kwlist,	&level, &ZstdCompressionDictType, &dict,
+		&ZstdCompressionParametersType, &params,
+		&writeChecksum, &writeContentSize, &writeDictID, &threads)) {
+		return -1;
+	}
+
+	if (level > ZSTD_maxCLevel()) {
+		PyErr_Format(PyExc_ValueError, "level must be less than %d",
+			ZSTD_maxCLevel() + 1);
+		return -1;
+	}
+
+	if (threads < 0) {
+		threads = cpu_count();
+	}
+
+	/* We create a ZSTD_CCtx for reuse among multiple operations to reduce the
+	   overhead of each compression operation. */
+	self->cctx = ZSTD_createCCtx();
+	if (!self->cctx) {
+		PyErr_NoMemory();
+		return -1;
+	}
+
+	/* TODO stuff the original parameters away somewhere so we can reset later. This
+	   will allow us to do things like automatically adjust cparams based on input
+	   size (assuming zstd isn't doing that internally). */
+
+	self->params = ZSTD_createCCtxParams();
+	if (!self->params) {
+		PyErr_NoMemory();
+		return -1;
+	}
+
+	if (params && writeChecksum) {
+		PyErr_SetString(PyExc_ValueError,
+			"cannot define compression_params and write_checksum");
+		return -1;
+	}
+
+	if (params && writeContentSize) {
+		PyErr_SetString(PyExc_ValueError,
+			"cannot define compression_params and write_content_size");
+		return -1;
+	}
+
+	if (params && writeDictID) {
+		PyErr_SetString(PyExc_ValueError,
+			"cannot define compression_params and write_dict_id");
+		return -1;
+	}
+
+	if (params && threads) {
+		PyErr_SetString(PyExc_ValueError,
+			"cannot define compression_params and threads");
+		return -1;
+	}
+
+	if (params) {
+		if (set_parameters(self->params, params)) {
+			return -1;
+		}
+	}
+	else {
+		if (set_parameter(self->params, ZSTD_c_compressionLevel, level)) {
+			return -1;
+		}
+
+		if (set_parameter(self->params, ZSTD_c_contentSizeFlag,
+			writeContentSize ? PyObject_IsTrue(writeContentSize) : 1)) {
+			return -1;
+		}
+
+		if (set_parameter(self->params, ZSTD_c_checksumFlag,
+			writeChecksum ? PyObject_IsTrue(writeChecksum) : 0)) {
+			return -1;
+		}
+
+		if (set_parameter(self->params, ZSTD_c_dictIDFlag,
+			writeDictID ? PyObject_IsTrue(writeDictID) : 1)) {
+			return -1;
+		}
+
+		if (threads) {
+			if (set_parameter(self->params, ZSTD_c_nbWorkers, threads)) {
+				return -1;
+			}
+		}
+	}
+
+	if (dict) {
+		self->dict = dict;
+		Py_INCREF(dict);
+	}
+
+    if (setup_cctx(self)) {
+        return -1;
+    }
+
+	return 0;
+}
+
+static void ZstdCompressor_dealloc(ZstdCompressor* self) {
+	if (self->cctx) {
+		ZSTD_freeCCtx(self->cctx);
+		self->cctx = NULL;
+	}
+
+	if (self->params) {
+		ZSTD_freeCCtxParams(self->params);
+		self->params = NULL;
+	}
+
+	Py_XDECREF(self->dict);
+	PyObject_Del(self);
+}
+
+PyDoc_STRVAR(ZstdCompressor_memory_size__doc__,
+"memory_size()\n"
+"\n"
+"Obtain the memory usage of this compressor, in bytes.\n"
+);
+
+static PyObject* ZstdCompressor_memory_size(ZstdCompressor* self) {
+	if (self->cctx) {
+		return PyLong_FromSize_t(ZSTD_sizeof_CCtx(self->cctx));
+	}
+	else {
+		PyErr_SetString(ZstdError, "no compressor context found; this should never happen");
+		return NULL;
+	}
+}
+
+PyDoc_STRVAR(ZstdCompressor_frame_progression__doc__,
+"frame_progression()\n"
+"\n"
+"Return information on how much work the compressor has done.\n"
+"\n"
+"Returns a 3-tuple of (ingested, consumed, produced).\n"
+);
+
+static PyObject* ZstdCompressor_frame_progression(ZstdCompressor* self) {
+	return frame_progression(self->cctx);
+}
+
+PyDoc_STRVAR(ZstdCompressor_copy_stream__doc__,
+"copy_stream(ifh, ofh[, size=0, read_size=default, write_size=default])\n"
+"compress data between streams\n"
+"\n"
+"Data will be read from ``ifh``, compressed, and written to ``ofh``.\n"
+"``ifh`` must have a ``read(size)`` method. ``ofh`` must have a ``write(data)``\n"
+"method.\n"
+"\n"
+"An optional ``size`` argument specifies the size of the source stream.\n"
+"If defined, compression parameters will be tuned based on the size.\n"
+"\n"
+"Optional arguments ``read_size`` and ``write_size`` define the chunk sizes\n"
+"of ``read()`` and ``write()`` operations, respectively. By default, they use\n"
+"the default compression stream input and output sizes, respectively.\n"
+);
+
+static PyObject* ZstdCompressor_copy_stream(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"ifh",
+		"ofh",
+		"size",
+		"read_size",
+		"write_size",
+		NULL
+	};
+
+	PyObject* source;
+	PyObject* dest;
+	unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+	size_t inSize = ZSTD_CStreamInSize();
+	size_t outSize = ZSTD_CStreamOutSize();
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	Py_ssize_t totalRead = 0;
+	Py_ssize_t totalWrite = 0;
+	char* readBuffer;
+	Py_ssize_t readSize;
+	PyObject* readResult = NULL;
+	PyObject* res = NULL;
+	size_t zresult;
+	PyObject* writeResult;
+	PyObject* totalReadPy;
+	PyObject* totalWritePy;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "OO|Kkk:copy_stream", kwlist,
+		&source, &dest, &sourceSize, &inSize, &outSize)) {
+		return NULL;
+	}
+
+	if (!PyObject_HasAttrString(source, "read")) {
+		PyErr_SetString(PyExc_ValueError, "first argument must have a read() method");
+		return NULL;
+	}
+
+	if (!PyObject_HasAttrString(dest, "write")) {
+		PyErr_SetString(PyExc_ValueError, "second argument must have a write() method");
+		return NULL;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source size: %s",
+			ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	/* Prevent free on uninitialized memory in finally. */
+	output.dst = PyMem_Malloc(outSize);
+	if (!output.dst) {
+		PyErr_NoMemory();
+		res = NULL;
+		goto finally;
+	}
+	output.size = outSize;
+	output.pos = 0;
+
+	input.src = NULL;
+	input.size = 0;
+	input.pos = 0;
+
+	while (1) {
+		/* Try to read from source stream. */
+		readResult = PyObject_CallMethod(source, "read", "n", inSize);
+		if (!readResult) {
+			PyErr_SetString(ZstdError, "could not read() from source");
+			goto finally;
+		}
+
+		PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+
+		/* If no data was read, we're at EOF. */
+		if (0 == readSize) {
+			break;
+		}
+
+		totalRead += readSize;
+
+		/* Send data to compressor */
+		input.src = readBuffer;
+		input.size = readSize;
+		input.pos = 0;
+
+		while (input.pos < input.size) {
+			Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_compressStream2(self->cctx, &output, &input, ZSTD_e_continue);
+			Py_END_ALLOW_THREADS
+
+			if (ZSTD_isError(zresult)) {
+				res = NULL;
+				PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+				goto finally;
+			}
+
+			if (output.pos) {
+#if PY_MAJOR_VERSION >= 3
+				writeResult = PyObject_CallMethod(dest, "write", "y#",
+#else
+				writeResult = PyObject_CallMethod(dest, "write", "s#",
+#endif
+					output.dst, output.pos);
+				Py_XDECREF(writeResult);
+				totalWrite += output.pos;
+				output.pos = 0;
+			}
+		}
+
+		Py_CLEAR(readResult);
+	}
+
+	/* We've finished reading. Now flush the compressor stream. */
+	assert(input.pos == input.size);
+
+	while (1) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(self->cctx, &output, &input, ZSTD_e_end);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "error ending compression stream: %s",
+				ZSTD_getErrorName(zresult));
+			res = NULL;
+			goto finally;
+		}
+
+		if (output.pos) {
+#if PY_MAJOR_VERSION >= 3
+			writeResult = PyObject_CallMethod(dest, "write", "y#",
+#else
+			writeResult = PyObject_CallMethod(dest, "write", "s#",
+#endif
+				output.dst, output.pos);
+			totalWrite += output.pos;
+			Py_XDECREF(writeResult);
+			output.pos = 0;
+		}
+
+		if (!zresult) {
+			break;
+		}
+	}
+
+	totalReadPy = PyLong_FromSsize_t(totalRead);
+	totalWritePy = PyLong_FromSsize_t(totalWrite);
+	res = PyTuple_Pack(2, totalReadPy, totalWritePy);
+	Py_DECREF(totalReadPy);
+	Py_DECREF(totalWritePy);
+
+finally:
+	if (output.dst) {
+		PyMem_Free(output.dst);
+	}
+
+	Py_XDECREF(readResult);
+
+	return res;
+}
+
+PyDoc_STRVAR(ZstdCompressor_stream_reader__doc__,
+"stream_reader(source, [size=0])\n"
+"\n"
+"Obtain an object that behaves like an I/O stream.\n"
+"\n"
+"The source object can be any object with a ``read(size)`` method\n"
+"or an object that conforms to the buffer protocol.\n"
+);
+
+static ZstdCompressionReader* ZstdCompressor_stream_reader(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"source",
+		"size",
+		"read_size",
+		NULL
+	};
+
+	PyObject* source;
+	unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+	size_t readSize = ZSTD_CStreamInSize();
+	ZstdCompressionReader* result = NULL;
+	size_t zresult;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|Kk:stream_reader", kwlist,
+		&source, &sourceSize, &readSize)) {
+		return NULL;
+	}
+
+	result = (ZstdCompressionReader*)PyObject_CallObject((PyObject*)&ZstdCompressionReaderType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	result->entered = 0;
+	result->closed = 0;
+
+	if (PyObject_HasAttrString(source, "read")) {
+		result->reader = source;
+		Py_INCREF(source);
+		result->readSize = readSize;
+	}
+	else if (1 == PyObject_CheckBuffer(source)) {
+		if (0 != PyObject_GetBuffer(source, &result->buffer, PyBUF_CONTIG_RO)) {
+			goto except;
+		}
+
+		assert(result->buffer.len >= 0);
+
+		sourceSize = result->buffer.len;
+	}
+	else {
+		PyErr_SetString(PyExc_TypeError,
+			"must pass an object with a read() method or that conforms to the buffer protocol");
+		goto except;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source source: %s",
+			ZSTD_getErrorName(zresult));
+		goto except;
+	}
+
+	result->compressor = self;
+	Py_INCREF(self);
+
+	return result;
+
+except:
+	Py_CLEAR(result);
+
+	return NULL;
+}
+
+PyDoc_STRVAR(ZstdCompressor_compress__doc__,
+"compress(data)\n"
+"\n"
+"Compress data in a single operation.\n"
+"\n"
+"This is the simplest mechanism to perform compression: simply pass in a\n"
+"value and get a compressed value back. It is almost the most prone to abuse.\n"
+"The input and output values must fit in memory, so passing in very large\n"
+"values can result in excessive memory usage. For this reason, one of the\n"
+"streaming based APIs is preferred for larger values.\n"
+);
+
+static PyObject* ZstdCompressor_compress(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	Py_buffer source;
+	size_t destSize;
+	PyObject* output = NULL;
+	size_t zresult;
+	ZSTD_outBuffer outBuffer;
+	ZSTD_inBuffer inBuffer;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|O:compress",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*|O:compress",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	destSize = ZSTD_compressBound(source.len);
+	output = PyBytes_FromStringAndSize(NULL, destSize);
+	if (!output) {
+		goto finally;
+	}
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, source.len);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source size: %s",
+			ZSTD_getErrorName(zresult));
+		Py_CLEAR(output);
+		goto finally;
+	}
+
+	inBuffer.src = source.buf;
+	inBuffer.size = source.len;
+	inBuffer.pos = 0;
+
+	outBuffer.dst = PyBytes_AsString(output);
+	outBuffer.size = destSize;
+	outBuffer.pos = 0;
+
+	Py_BEGIN_ALLOW_THREADS
+	/* By avoiding ZSTD_compress(), we don't necessarily write out content
+		size. This means the argument to ZstdCompressor to control frame
+		parameters is honored. */
+	zresult = ZSTD_compressStream2(self->cctx, &outBuffer, &inBuffer, ZSTD_e_end);
+	Py_END_ALLOW_THREADS
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "cannot compress: %s", ZSTD_getErrorName(zresult));
+		Py_CLEAR(output);
+		goto finally;
+	}
+	else if (zresult) {
+		PyErr_SetString(ZstdError, "unexpected partial frame flush");
+		Py_CLEAR(output);
+		goto finally;
+	}
+
+	Py_SIZE(output) = outBuffer.pos;
+
+finally:
+	PyBuffer_Release(&source);
+	return output;
+}
+
+PyDoc_STRVAR(ZstdCompressionObj__doc__,
+"compressobj()\n"
+"\n"
+"Return an object exposing ``compress(data)`` and ``flush()`` methods.\n"
+"\n"
+"The returned object exposes an API similar to ``zlib.compressobj`` and\n"
+"``bz2.BZ2Compressor`` so that callers can swap in the zstd compressor\n"
+"without changing how compression is performed.\n"
+);
+
+static ZstdCompressionObj* ZstdCompressor_compressobj(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"size",
+		NULL
+	};
+
+	unsigned long long inSize = ZSTD_CONTENTSIZE_UNKNOWN;
+	size_t outSize = ZSTD_CStreamOutSize();
+	ZstdCompressionObj* result = NULL;
+	size_t zresult;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|K:compressobj", kwlist, &inSize)) {
+		return NULL;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, inSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source size: %s",
+			ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	result = (ZstdCompressionObj*)PyObject_CallObject((PyObject*)&ZstdCompressionObjType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	result->output.dst = PyMem_Malloc(outSize);
+	if (!result->output.dst) {
+		PyErr_NoMemory();
+		Py_DECREF(result);
+		return NULL;
+	}
+	result->output.size = outSize;
+	result->compressor = self;
+	Py_INCREF(result->compressor);
+
+	return result;
+}
+
+PyDoc_STRVAR(ZstdCompressor_read_to_iter__doc__,
+"read_to_iter(reader, [size=0, read_size=default, write_size=default])\n"
+"Read uncompressed data from a reader and return an iterator\n"
+"\n"
+"Returns an iterator of compressed data produced from reading from ``reader``.\n"
+"\n"
+"Uncompressed data will be obtained from ``reader`` by calling the\n"
+"``read(size)`` method of it. The source data will be streamed into a\n"
+"compressor. As compressed data is available, it will be exposed to the\n"
+"iterator.\n"
+"\n"
+"Data is read from the source in chunks of ``read_size``. Compressed chunks\n"
+"are at most ``write_size`` bytes. Both values default to the zstd input and\n"
+"and output defaults, respectively.\n"
+"\n"
+"The caller is partially in control of how fast data is fed into the\n"
+"compressor by how it consumes the returned iterator. The compressor will\n"
+"not consume from the reader unless the caller consumes from the iterator.\n"
+);
+
+static ZstdCompressorIterator* ZstdCompressor_read_to_iter(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"reader",
+		"size",
+		"read_size",
+		"write_size",
+		NULL
+	};
+
+	PyObject* reader;
+	unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+	size_t inSize = ZSTD_CStreamInSize();
+	size_t outSize = ZSTD_CStreamOutSize();
+	ZstdCompressorIterator* result;
+	size_t zresult;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|Kkk:read_to_iter", kwlist,
+		&reader, &sourceSize, &inSize, &outSize)) {
+		return NULL;
+	}
+
+	result = (ZstdCompressorIterator*)PyObject_CallObject((PyObject*)&ZstdCompressorIteratorType, NULL);
+	if (!result) {
+		return NULL;
+	}
+	if (PyObject_HasAttrString(reader, "read")) {
+		result->reader = reader;
+		Py_INCREF(result->reader);
+	}
+	else if (1 == PyObject_CheckBuffer(reader)) {
+		if (0 != PyObject_GetBuffer(reader, &result->buffer, PyBUF_CONTIG_RO)) {
+			goto except;
+		}
+
+		sourceSize = result->buffer.len;
+	}
+	else {
+		PyErr_SetString(PyExc_ValueError,
+			"must pass an object with a read() method or conforms to buffer protocol");
+		goto except;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source size: %s",
+			ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	result->compressor = self;
+	Py_INCREF(result->compressor);
+
+	result->inSize = inSize;
+	result->outSize = outSize;
+
+	result->output.dst = PyMem_Malloc(outSize);
+	if (!result->output.dst) {
+		PyErr_NoMemory();
+		goto except;
+	}
+	result->output.size = outSize;
+
+	goto finally;
+
+except:
+	Py_CLEAR(result);
+
+finally:
+	return result;
+}
+
+PyDoc_STRVAR(ZstdCompressor_stream_writer___doc__,
+"Create a context manager to write compressed data to an object.\n"
+"\n"
+"The passed object must have a ``write()`` method.\n"
+"\n"
+"The caller feeds input data to the object by calling ``compress(data)``.\n"
+"Compressed data is written to the argument given to this function.\n"
+"\n"
+"The function takes an optional ``size`` argument indicating the total size\n"
+"of the eventual input. If specified, the size will influence compression\n"
+"parameter tuning and could result in the size being written into the\n"
+"header of the compressed data.\n"
+"\n"
+"An optional ``write_size`` argument is also accepted. It defines the maximum\n"
+"byte size of chunks fed to ``write()``. By default, it uses the zstd default\n"
+"for a compressor output stream.\n"
+);
+
+static ZstdCompressionWriter* ZstdCompressor_stream_writer(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"writer",
+		"size",
+		"write_size",
+		"write_return_read",
+		NULL
+	};
+
+	PyObject* writer;
+	ZstdCompressionWriter* result;
+	size_t zresult;
+	unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+	size_t outSize = ZSTD_CStreamOutSize();
+	PyObject* writeReturnRead = NULL;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|KkO:stream_writer", kwlist,
+		&writer, &sourceSize, &outSize, &writeReturnRead)) {
+		return NULL;
+	}
+
+	if (!PyObject_HasAttrString(writer, "write")) {
+		PyErr_SetString(PyExc_ValueError, "must pass an object with a write() method");
+		return NULL;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source size: %s",
+			ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	result = (ZstdCompressionWriter*)PyObject_CallObject((PyObject*)&ZstdCompressionWriterType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	result->entered = 0;
+	result->closed = 0;
+
+	result->output.dst = PyMem_Malloc(outSize);
+	if (!result->output.dst) {
+		Py_DECREF(result);
+		return (ZstdCompressionWriter*)PyErr_NoMemory();
+	}
+
+	result->output.pos = 0;
+	result->output.size = outSize;
+
+	result->compressor = self;
+	Py_INCREF(result->compressor);
+
+	result->writer = writer;
+	Py_INCREF(result->writer);
+
+	result->outSize = outSize;
+	result->bytesCompressed = 0;
+	result->writeReturnRead = writeReturnRead ? PyObject_IsTrue(writeReturnRead) : 0;
+
+	return result;
+}
+
+PyDoc_STRVAR(ZstdCompressor_chunker__doc__,
+"Create an object for iterative compressing to same-sized chunks.\n"
+);
+
+static ZstdCompressionChunker* ZstdCompressor_chunker(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"size",
+		"chunk_size",
+		NULL
+	};
+
+	unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+	size_t chunkSize = ZSTD_CStreamOutSize();
+	ZstdCompressionChunker* chunker;
+	size_t zresult;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|Kk:chunker", kwlist,
+		&sourceSize, &chunkSize)) {
+		return NULL;
+	}
+
+	ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+	zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "error setting source size: %s",
+			ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	chunker = (ZstdCompressionChunker*)PyObject_CallObject((PyObject*)&ZstdCompressionChunkerType, NULL);
+	if (!chunker) {
+		return NULL;
+	}
+
+	chunker->output.dst = PyMem_Malloc(chunkSize);
+	if (!chunker->output.dst) {
+		PyErr_NoMemory();
+		Py_DECREF(chunker);
+		return NULL;
+	}
+	chunker->output.size = chunkSize;
+	chunker->output.pos = 0;
+
+	chunker->compressor = self;
+	Py_INCREF(chunker->compressor);
+
+	chunker->chunkSize = chunkSize;
+
+	return chunker;
+}
+
+typedef struct {
+	void* sourceData;
+	size_t sourceSize;
+} DataSource;
+
+typedef struct {
+	DataSource* sources;
+	Py_ssize_t sourcesSize;
+	unsigned long long totalSourceSize;
+} DataSources;
+
+typedef struct {
+	void* dest;
+	Py_ssize_t destSize;
+	BufferSegment* segments;
+	Py_ssize_t segmentsSize;
+} DestBuffer;
+
+typedef enum {
+	WorkerError_none = 0,
+	WorkerError_zstd = 1,
+	WorkerError_no_memory = 2,
+	WorkerError_nospace = 3,
+} WorkerError;
+
+/**
+ * Holds state for an individual worker performing multi_compress_to_buffer work.
+ */
+typedef struct {
+	/* Used for compression. */
+	ZSTD_CCtx* cctx;
+
+	/* What to compress. */
+	DataSource* sources;
+	Py_ssize_t sourcesSize;
+	Py_ssize_t startOffset;
+	Py_ssize_t endOffset;
+	unsigned long long totalSourceSize;
+
+	/* Result storage. */
+	DestBuffer* destBuffers;
+	Py_ssize_t destCount;
+
+	/* Error tracking. */
+	WorkerError error;
+	size_t zresult;
+	Py_ssize_t errorOffset;
+} WorkerState;
+
+static void compress_worker(WorkerState* state) {
+	Py_ssize_t inputOffset = state->startOffset;
+	Py_ssize_t remainingItems = state->endOffset - state->startOffset + 1;
+	Py_ssize_t currentBufferStartOffset = state->startOffset;
+	size_t zresult;
+	void* newDest;
+	size_t allocationSize;
+	size_t boundSize;
+	Py_ssize_t destOffset = 0;
+	DataSource* sources = state->sources;
+	DestBuffer* destBuffer;
+
+	assert(!state->destBuffers);
+	assert(0 == state->destCount);
+
+	/*
+	 * The total size of the compressed data is unknown until we actually
+	 * compress data. That means we can't pre-allocate the exact size we need.
+	 *
+	 * There is a cost to every allocation and reallocation. So, it is in our
+	 * interest to minimize the number of allocations.
+	 *
+	 * There is also a cost to too few allocations. If allocations are too
+	 * large they may fail. If buffers are shared and all inputs become
+	 * irrelevant at different lifetimes, then a reference to one segment
+	 * in the buffer will keep the entire buffer alive. This leads to excessive
+	 * memory usage.
+	 *
+	 * Our current strategy is to assume a compression ratio of 16:1 and
+	 * allocate buffers of that size, rounded up to the nearest power of 2
+	 * (because computers like round numbers). That ratio is greater than what
+	 * most inputs achieve. This is by design: we don't want to over-allocate.
+	 * But we don't want to under-allocate and lead to too many buffers either.
+	 */
+
+	state->destCount = 1;
+
+	state->destBuffers = calloc(1, sizeof(DestBuffer));
+	if (NULL == state->destBuffers) {
+		state->error = WorkerError_no_memory;
+		return;
+	}
+
+	destBuffer = &state->destBuffers[state->destCount - 1];
+
+	/*
+	 * Rather than track bounds and grow the segments buffer, allocate space
+	 * to hold remaining items then truncate when we're done with it.
+	 */
+	destBuffer->segments = calloc(remainingItems, sizeof(BufferSegment));
+	if (NULL == destBuffer->segments) {
+		state->error = WorkerError_no_memory;
+		return;
+	}
+
+	destBuffer->segmentsSize = remainingItems;
+
+	assert(state->totalSourceSize <= SIZE_MAX);
+	allocationSize = roundpow2((size_t)state->totalSourceSize >> 4);
+
+	/* If the maximum size of the output is larger than that, round up. */
+	boundSize = ZSTD_compressBound(sources[inputOffset].sourceSize);
+
+	if (boundSize > allocationSize) {
+		allocationSize = roundpow2(boundSize);
+	}
+
+	destBuffer->dest = malloc(allocationSize);
+	if (NULL == destBuffer->dest) {
+		state->error = WorkerError_no_memory;
+		return;
+	}
+
+	destBuffer->destSize = allocationSize;
+
+	for (inputOffset = state->startOffset; inputOffset <= state->endOffset; inputOffset++) {
+		void* source = sources[inputOffset].sourceData;
+		size_t sourceSize = sources[inputOffset].sourceSize;
+		size_t destAvailable;
+		void* dest;
+		ZSTD_outBuffer opOutBuffer;
+		ZSTD_inBuffer opInBuffer;
+
+		destAvailable = destBuffer->destSize - destOffset;
+		boundSize = ZSTD_compressBound(sourceSize);
+
+		/*
+		 * Not enough space in current buffer to hold largest compressed output.
+		 * So allocate and switch to a new output buffer.
+		 */
+		if (boundSize > destAvailable) {
+			/*
+			 * The downsizing of the existing buffer is optional. It should be cheap
+			 * (unlike growing). So we just do it.
+			 */
+			if (destAvailable) {
+				newDest = realloc(destBuffer->dest, destOffset);
+				if (NULL == newDest) {
+					state->error = WorkerError_no_memory;
+					return;
+				}
+
+				destBuffer->dest = newDest;
+				destBuffer->destSize = destOffset;
+			}
+
+			/* Truncate segments buffer. */
+			newDest = realloc(destBuffer->segments,
+				(inputOffset - currentBufferStartOffset + 1) * sizeof(BufferSegment));
+			if (NULL == newDest) {
+				state->error = WorkerError_no_memory;
+				return;
+			}
+
+			destBuffer->segments = newDest;
+			destBuffer->segmentsSize = inputOffset - currentBufferStartOffset;
+
+			/* Grow space for new struct. */
+			/* TODO consider over-allocating so we don't do this every time. */
+			newDest = realloc(state->destBuffers, (state->destCount + 1) * sizeof(DestBuffer));
+			if (NULL == newDest) {
+				state->error = WorkerError_no_memory;
+				return;
+			}
+
+			state->destBuffers = newDest;
+			state->destCount++;
+
+			destBuffer = &state->destBuffers[state->destCount - 1];
+
+			/* Don't take any chances with non-NULL pointers. */
+			memset(destBuffer, 0, sizeof(DestBuffer));
+
+			/**
+			 * We could dynamically update allocation size based on work done so far.
+			 * For now, keep is simple.
+			 */
+			assert(state->totalSourceSize <= SIZE_MAX);
+			allocationSize = roundpow2((size_t)state->totalSourceSize >> 4);
+
+			if (boundSize > allocationSize) {
+				allocationSize = roundpow2(boundSize);
+			}
+
+			destBuffer->dest = malloc(allocationSize);
+			if (NULL == destBuffer->dest) {
+				state->error = WorkerError_no_memory;
+				return;
+			}
+
+			destBuffer->destSize = allocationSize;
+			destAvailable = allocationSize;
+			destOffset = 0;
+
+			destBuffer->segments = calloc(remainingItems, sizeof(BufferSegment));
+			if (NULL == destBuffer->segments) {
+				state->error = WorkerError_no_memory;
+				return;
+			}
+
+			destBuffer->segmentsSize = remainingItems;
+			currentBufferStartOffset = inputOffset;
+		}
+
+		dest = (char*)destBuffer->dest + destOffset;
+
+		opInBuffer.src = source;
+		opInBuffer.size = sourceSize;
+		opInBuffer.pos = 0;
+
+		opOutBuffer.dst = dest;
+		opOutBuffer.size = destAvailable;
+		opOutBuffer.pos = 0;
+
+		zresult = ZSTD_CCtx_setPledgedSrcSize(state->cctx, sourceSize);
+		if (ZSTD_isError(zresult)) {
+			state->error = WorkerError_zstd;
+			state->zresult = zresult;
+			state->errorOffset = inputOffset;
+			break;
+		}
+
+		zresult = ZSTD_compressStream2(state->cctx, &opOutBuffer, &opInBuffer, ZSTD_e_end);
+		if (ZSTD_isError(zresult)) {
+			state->error = WorkerError_zstd;
+			state->zresult = zresult;
+			state->errorOffset = inputOffset;
+			break;
+		}
+		else if (zresult) {
+			state->error = WorkerError_nospace;
+			state->errorOffset = inputOffset;
+			break;
+		}
+
+		destBuffer->segments[inputOffset - currentBufferStartOffset].offset = destOffset;
+		destBuffer->segments[inputOffset - currentBufferStartOffset].length = opOutBuffer.pos;
+
+		destOffset += opOutBuffer.pos;
+		remainingItems--;
+	}
+
+	if (destBuffer->destSize > destOffset) {
+		newDest = realloc(destBuffer->dest, destOffset);
+		if (NULL == newDest) {
+			state->error = WorkerError_no_memory;
+			return;
+		}
+
+		destBuffer->dest = newDest;
+		destBuffer->destSize = destOffset;
+	}
+}
+
+ZstdBufferWithSegmentsCollection* compress_from_datasources(ZstdCompressor* compressor,
+	DataSources* sources, Py_ssize_t threadCount) {
+	unsigned long long bytesPerWorker;
+	POOL_ctx* pool = NULL;
+	WorkerState* workerStates = NULL;
+	Py_ssize_t i;
+	unsigned long long workerBytes = 0;
+	Py_ssize_t workerStartOffset = 0;
+	Py_ssize_t currentThread = 0;
+	int errored = 0;
+	Py_ssize_t segmentsCount = 0;
+	Py_ssize_t segmentIndex;
+	PyObject* segmentsArg = NULL;
+	ZstdBufferWithSegments* buffer;
+	ZstdBufferWithSegmentsCollection* result = NULL;
+
+	assert(sources->sourcesSize > 0);
+	assert(sources->totalSourceSize > 0);
+	assert(threadCount >= 1);
+
+	/* More threads than inputs makes no sense. */
+	threadCount = sources->sourcesSize < threadCount ? sources->sourcesSize
+													 : threadCount;
+
+	/* TODO lower thread count when input size is too small and threads would add
+	overhead. */
+
+	workerStates = PyMem_Malloc(threadCount * sizeof(WorkerState));
+	if (NULL == workerStates) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+
+	memset(workerStates, 0, threadCount * sizeof(WorkerState));
+
+	if (threadCount > 1) {
+		pool = POOL_create(threadCount, 1);
+		if (NULL == pool) {
+			PyErr_SetString(ZstdError, "could not initialize zstd thread pool");
+			goto finally;
+		}
+	}
+
+	bytesPerWorker = sources->totalSourceSize / threadCount;
+
+	for (i = 0; i < threadCount; i++) {
+		size_t zresult;
+
+		workerStates[i].cctx = ZSTD_createCCtx();
+		if (!workerStates[i].cctx) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		zresult = ZSTD_CCtx_setParametersUsingCCtxParams(workerStates[i].cctx,
+			compressor->params);
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "could not set compression parameters: %s",
+				ZSTD_getErrorName(zresult));
+			goto finally;
+		}
+
+		if (compressor->dict) {
+			if (compressor->dict->cdict) {
+				zresult = ZSTD_CCtx_refCDict(workerStates[i].cctx, compressor->dict->cdict);
+			}
+			else {
+				zresult = ZSTD_CCtx_loadDictionary_advanced(
+					workerStates[i].cctx,
+					compressor->dict->dictData,
+					compressor->dict->dictSize,
+					ZSTD_dlm_byRef,
+					compressor->dict->dictType);
+			}
+
+			if (ZSTD_isError(zresult)) {
+				PyErr_Format(ZstdError, "could not load compression dictionary: %s",
+					ZSTD_getErrorName(zresult));
+				goto finally;
+			}
+
+		}
+
+		workerStates[i].sources = sources->sources;
+		workerStates[i].sourcesSize = sources->sourcesSize;
+	}
+
+	Py_BEGIN_ALLOW_THREADS
+	for (i = 0; i < sources->sourcesSize; i++) {
+		workerBytes += sources->sources[i].sourceSize;
+
+		/*
+		 * The last worker/thread needs to handle all remaining work. Don't
+		 * trigger it prematurely. Defer to the block outside of the loop
+		 * to run the last worker/thread. But do still process this loop
+		 * so workerBytes is correct.
+		 */
+		if (currentThread == threadCount - 1) {
+			continue;
+		}
+
+		if (workerBytes >= bytesPerWorker) {
+			assert(currentThread < threadCount);
+			workerStates[currentThread].totalSourceSize = workerBytes;
+			workerStates[currentThread].startOffset = workerStartOffset;
+			workerStates[currentThread].endOffset = i;
+
+			if (threadCount > 1) {
+				POOL_add(pool, (POOL_function)compress_worker, &workerStates[currentThread]);
+			}
+			else {
+				compress_worker(&workerStates[currentThread]);
+			}
+
+			currentThread++;
+			workerStartOffset = i + 1;
+			workerBytes = 0;
+		}
+	}
+
+	if (workerBytes) {
+		assert(currentThread < threadCount);
+		workerStates[currentThread].totalSourceSize = workerBytes;
+		workerStates[currentThread].startOffset = workerStartOffset;
+		workerStates[currentThread].endOffset = sources->sourcesSize - 1;
+
+		if (threadCount > 1) {
+			POOL_add(pool, (POOL_function)compress_worker, &workerStates[currentThread]);
+		}
+		else {
+			compress_worker(&workerStates[currentThread]);
+		}
+	}
+
+	if (threadCount > 1) {
+		POOL_free(pool);
+		pool = NULL;
+	}
+
+	Py_END_ALLOW_THREADS
+
+	for (i = 0; i < threadCount; i++) {
+		switch (workerStates[i].error) {
+		case WorkerError_no_memory:
+			PyErr_NoMemory();
+			errored = 1;
+			break;
+
+		case WorkerError_zstd:
+			PyErr_Format(ZstdError, "error compressing item %zd: %s",
+				workerStates[i].errorOffset, ZSTD_getErrorName(workerStates[i].zresult));
+			errored = 1;
+			break;
+
+		case WorkerError_nospace:
+			PyErr_Format(ZstdError, "error compressing item %zd: not enough space in output",
+				workerStates[i].errorOffset);
+			errored = 1;
+			break;
+
+		default:
+			;
+		}
+
+		if (errored) {
+			break;
+		}
+
+	}
+
+	if (errored) {
+		goto finally;
+	}
+
+	segmentsCount = 0;
+	for (i = 0; i < threadCount; i++) {
+		WorkerState* state = &workerStates[i];
+		segmentsCount += state->destCount;
+	}
+
+	segmentsArg = PyTuple_New(segmentsCount);
+	if (NULL == segmentsArg) {
+		goto finally;
+	}
+
+	segmentIndex = 0;
+
+	for (i = 0; i < threadCount; i++) {
+		Py_ssize_t j;
+		WorkerState* state = &workerStates[i];
+
+		for (j = 0; j < state->destCount; j++) {
+			DestBuffer* destBuffer = &state->destBuffers[j];
+			buffer = BufferWithSegments_FromMemory(destBuffer->dest, destBuffer->destSize,
+				destBuffer->segments, destBuffer->segmentsSize);
+
+			if (NULL == buffer) {
+				goto finally;
+			}
+
+			/* Tell instance to use free() instsead of PyMem_Free(). */
+			buffer->useFree = 1;
+
+			/*
+			 * BufferWithSegments_FromMemory takes ownership of the backing memory.
+			 * Unset it here so it doesn't get freed below.
+			 */
+			destBuffer->dest = NULL;
+			destBuffer->segments = NULL;
+
+			PyTuple_SET_ITEM(segmentsArg, segmentIndex++, (PyObject*)buffer);
+		}
+	}
+
+	result = (ZstdBufferWithSegmentsCollection*)PyObject_CallObject(
+		(PyObject*)&ZstdBufferWithSegmentsCollectionType, segmentsArg);
+
+finally:
+	Py_CLEAR(segmentsArg);
+
+	if (pool) {
+		POOL_free(pool);
+	}
+
+	if (workerStates) {
+		Py_ssize_t j;
+
+		for (i = 0; i < threadCount; i++) {
+			WorkerState state = workerStates[i];
+
+			if (state.cctx) {
+				ZSTD_freeCCtx(state.cctx);
+			}
+
+			/* malloc() is used in worker thread. */
+
+			for (j = 0; j < state.destCount; j++) {
+				if (state.destBuffers) {
+					free(state.destBuffers[j].dest);
+					free(state.destBuffers[j].segments);
+				}
+			}
+
+
+			free(state.destBuffers);
+		}
+
+		PyMem_Free(workerStates);
+	}
+
+	return result;
+}
+
+PyDoc_STRVAR(ZstdCompressor_multi_compress_to_buffer__doc__,
+"Compress multiple pieces of data as a single operation\n"
+"\n"
+"Receives a ``BufferWithSegmentsCollection``, a ``BufferWithSegments``, or\n"
+"a list of bytes like objects holding data to compress.\n"
+"\n"
+"Returns a ``BufferWithSegmentsCollection`` holding compressed data.\n"
+"\n"
+"This function is optimized to perform multiple compression operations as\n"
+"as possible with as little overhead as possbile.\n"
+);
+
+static ZstdBufferWithSegmentsCollection* ZstdCompressor_multi_compress_to_buffer(ZstdCompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		"threads",
+		NULL
+	};
+
+	PyObject* data;
+	int threads = 0;
+	Py_buffer* dataBuffers = NULL;
+	DataSources sources;
+	Py_ssize_t i;
+	Py_ssize_t sourceCount = 0;
+	ZstdBufferWithSegmentsCollection* result = NULL;
+
+	memset(&sources, 0, sizeof(sources));
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|i:multi_compress_to_buffer", kwlist,
+		&data, &threads)) {
+		return NULL;
+	}
+
+	if (threads < 0) {
+		threads = cpu_count();
+	}
+
+	if (threads < 2) {
+		threads = 1;
+	}
+
+	if (PyObject_TypeCheck(data, &ZstdBufferWithSegmentsType)) {
+		ZstdBufferWithSegments* buffer = (ZstdBufferWithSegments*)data;
+
+		sources.sources = PyMem_Malloc(buffer->segmentCount * sizeof(DataSource));
+		if (NULL == sources.sources) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		for (i = 0; i < buffer->segmentCount; i++) {
+			if (buffer->segments[i].length > SIZE_MAX) {
+				PyErr_Format(PyExc_ValueError,
+					"buffer segment %zd is too large for this platform", i);
+				goto finally;
+			}
+
+			sources.sources[i].sourceData = (char*)buffer->data + buffer->segments[i].offset;
+			sources.sources[i].sourceSize = (size_t)buffer->segments[i].length;
+			sources.totalSourceSize += buffer->segments[i].length;
+		}
+
+		sources.sourcesSize = buffer->segmentCount;
+	}
+	else if (PyObject_TypeCheck(data, &ZstdBufferWithSegmentsCollectionType)) {
+		Py_ssize_t j;
+		Py_ssize_t offset = 0;
+		ZstdBufferWithSegments* buffer;
+		ZstdBufferWithSegmentsCollection* collection = (ZstdBufferWithSegmentsCollection*)data;
+
+		sourceCount = BufferWithSegmentsCollection_length(collection);
+
+		sources.sources = PyMem_Malloc(sourceCount * sizeof(DataSource));
+		if (NULL == sources.sources) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		for (i = 0; i < collection->bufferCount; i++) {
+			buffer = collection->buffers[i];
+
+			for (j = 0; j < buffer->segmentCount; j++) {
+				if (buffer->segments[j].length > SIZE_MAX) {
+					PyErr_Format(PyExc_ValueError,
+						"buffer segment %zd in buffer %zd is too large for this platform",
+						j, i);
+					goto finally;
+				}
+
+				sources.sources[offset].sourceData = (char*)buffer->data + buffer->segments[j].offset;
+				sources.sources[offset].sourceSize = (size_t)buffer->segments[j].length;
+				sources.totalSourceSize += buffer->segments[j].length;
+
+				offset++;
+			}
+		}
+
+		sources.sourcesSize = sourceCount;
+	}
+	else if (PyList_Check(data)) {
+		sourceCount = PyList_GET_SIZE(data);
+
+		sources.sources = PyMem_Malloc(sourceCount * sizeof(DataSource));
+		if (NULL == sources.sources) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		dataBuffers = PyMem_Malloc(sourceCount * sizeof(Py_buffer));
+		if (NULL == dataBuffers) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		memset(dataBuffers, 0, sourceCount * sizeof(Py_buffer));
+
+		for (i = 0; i < sourceCount; i++) {
+			if (0 != PyObject_GetBuffer(PyList_GET_ITEM(data, i),
+				&dataBuffers[i], PyBUF_CONTIG_RO)) {
+				PyErr_Clear();
+				PyErr_Format(PyExc_TypeError, "item %zd not a bytes like object", i);
+				goto finally;
+			}
+
+			sources.sources[i].sourceData = dataBuffers[i].buf;
+			sources.sources[i].sourceSize = dataBuffers[i].len;
+			sources.totalSourceSize += dataBuffers[i].len;
+		}
+
+		sources.sourcesSize = sourceCount;
+	}
+	else {
+		PyErr_SetString(PyExc_TypeError, "argument must be list of BufferWithSegments");
+		goto finally;
+	}
+
+	if (0 == sources.sourcesSize) {
+		PyErr_SetString(PyExc_ValueError, "no source elements found");
+		goto finally;
+	}
+
+	if (0 == sources.totalSourceSize) {
+		PyErr_SetString(PyExc_ValueError, "source elements are empty");
+		goto finally;
+	}
+
+	if (sources.totalSourceSize > SIZE_MAX) {
+		PyErr_SetString(PyExc_ValueError, "sources are too large for this platform");
+		goto finally;
+	}
+
+	result = compress_from_datasources(self, &sources, threads);
+
+finally:
+	PyMem_Free(sources.sources);
+
+	if (dataBuffers) {
+		for (i = 0; i < sourceCount; i++) {
+			PyBuffer_Release(&dataBuffers[i]);
+		}
+
+		PyMem_Free(dataBuffers);
+	}
+
+	return result;
+}
+
+static PyMethodDef ZstdCompressor_methods[] = {
+	{ "chunker", (PyCFunction)ZstdCompressor_chunker,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_chunker__doc__ },
+	{ "compress", (PyCFunction)ZstdCompressor_compress,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_compress__doc__ },
+	{ "compressobj", (PyCFunction)ZstdCompressor_compressobj,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressionObj__doc__ },
+	{ "copy_stream", (PyCFunction)ZstdCompressor_copy_stream,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_copy_stream__doc__ },
+	{ "stream_reader", (PyCFunction)ZstdCompressor_stream_reader,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_stream_reader__doc__ },
+	{ "stream_writer", (PyCFunction)ZstdCompressor_stream_writer,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_stream_writer___doc__ },
+	{ "read_to_iter", (PyCFunction)ZstdCompressor_read_to_iter,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_read_to_iter__doc__ },
+	/* TODO Remove deprecated API */
+	{ "read_from", (PyCFunction)ZstdCompressor_read_to_iter,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_read_to_iter__doc__ },
+	/* TODO remove deprecated API */
+	{ "write_to", (PyCFunction)ZstdCompressor_stream_writer,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_stream_writer___doc__ },
+	{ "multi_compress_to_buffer", (PyCFunction)ZstdCompressor_multi_compress_to_buffer,
+	METH_VARARGS | METH_KEYWORDS, ZstdCompressor_multi_compress_to_buffer__doc__ },
+	{ "memory_size", (PyCFunction)ZstdCompressor_memory_size,
+	METH_NOARGS, ZstdCompressor_memory_size__doc__ },
+	{ "frame_progression", (PyCFunction)ZstdCompressor_frame_progression,
+	METH_NOARGS, ZstdCompressor_frame_progression__doc__ },
+	{ NULL, NULL }
+};
+
+PyTypeObject ZstdCompressorType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressor",         /* tp_name */
+	sizeof(ZstdCompressor),        /* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)ZstdCompressor_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	0,                              /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressor__doc__,          /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	ZstdCompressor_methods,         /* tp_methods */
+	0,                              /* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	(initproc)ZstdCompressor_init,  /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void compressor_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdCompressorType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressorType) < 0) {
+		return;
+	}
+
+	Py_INCREF((PyObject*)&ZstdCompressorType);
+	PyModule_AddObject(mod, "ZstdCompressor",
+		(PyObject*)&ZstdCompressorType);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/compressoriterator.c b/contrib/python/zstandard/py2/c-ext/compressoriterator.c
new file mode 100644
index 00000000000..24e31c5a4be
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/compressoriterator.c
@@ -0,0 +1,235 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+#define min(a, b) (((a) < (b)) ? (a) : (b))
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(ZstdCompressorIterator__doc__,
+"Represents an iterator of compressed data.\n"
+);
+
+static void ZstdCompressorIterator_dealloc(ZstdCompressorIterator* self) {
+	Py_XDECREF(self->readResult);
+	Py_XDECREF(self->compressor);
+	Py_XDECREF(self->reader);
+
+	if (self->buffer.buf) {
+		PyBuffer_Release(&self->buffer);
+		memset(&self->buffer, 0, sizeof(self->buffer));
+	}
+
+	if (self->output.dst) {
+		PyMem_Free(self->output.dst);
+		self->output.dst = NULL;
+	}
+
+	PyObject_Del(self);
+}
+
+static PyObject* ZstdCompressorIterator_iter(PyObject* self) {
+	Py_INCREF(self);
+	return self;
+}
+
+static PyObject* ZstdCompressorIterator_iternext(ZstdCompressorIterator* self) {
+	size_t zresult;
+	PyObject* readResult = NULL;
+	PyObject* chunk;
+	char* readBuffer;
+	Py_ssize_t readSize = 0;
+	Py_ssize_t bufferRemaining;
+
+	if (self->finishedOutput) {
+		PyErr_SetString(PyExc_StopIteration, "output flushed");
+		return NULL;
+	}
+
+feedcompressor:
+
+	/* If we have data left in the input, consume it. */
+	if (self->input.pos < self->input.size) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+			&self->input, ZSTD_e_continue);
+		Py_END_ALLOW_THREADS
+
+		/* Release the Python object holding the input buffer. */
+		if (self->input.pos == self->input.size) {
+			self->input.src = NULL;
+			self->input.pos = 0;
+			self->input.size = 0;
+			Py_DECREF(self->readResult);
+			self->readResult = NULL;
+		}
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+			return NULL;
+		}
+
+		/* If it produced output data, emit it. */
+		if (self->output.pos) {
+			chunk = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+			self->output.pos = 0;
+			return chunk;
+		}
+	}
+
+	/* We should never have output data sitting around after a previous call. */
+	assert(self->output.pos == 0);
+
+	/* The code above should have either emitted a chunk and returned or consumed
+	the entire input buffer. So the state of the input buffer is not
+	relevant. */
+	if (!self->finishedInput) {
+		if (self->reader) {
+			readResult = PyObject_CallMethod(self->reader, "read", "I", self->inSize);
+			if (!readResult) {
+				PyErr_SetString(ZstdError, "could not read() from source");
+				return NULL;
+			}
+
+			PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+		}
+		else {
+			assert(self->buffer.buf);
+
+			/* Only support contiguous C arrays. */
+			assert(self->buffer.strides == NULL && self->buffer.suboffsets == NULL);
+			assert(self->buffer.itemsize == 1);
+
+			readBuffer = (char*)self->buffer.buf + self->bufferOffset;
+			bufferRemaining = self->buffer.len - self->bufferOffset;
+			readSize = min(bufferRemaining, (Py_ssize_t)self->inSize);
+			self->bufferOffset += readSize;
+		}
+
+		if (0 == readSize) {
+			Py_XDECREF(readResult);
+			self->finishedInput = 1;
+		}
+		else {
+			self->readResult = readResult;
+		}
+	}
+
+	/* EOF */
+	if (0 == readSize) {
+		self->input.src = NULL;
+		self->input.size = 0;
+		self->input.pos = 0;
+
+		zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+			&self->input, ZSTD_e_end);
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "error ending compression stream: %s",
+				ZSTD_getErrorName(zresult));
+			return NULL;
+		}
+
+		assert(self->output.pos);
+
+		if (0 == zresult) {
+			self->finishedOutput = 1;
+		}
+
+		chunk = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+		self->output.pos = 0;
+		return chunk;
+	}
+
+	/* New data from reader. Feed into compressor. */
+	self->input.src = readBuffer;
+	self->input.size = readSize;
+	self->input.pos = 0;
+
+	Py_BEGIN_ALLOW_THREADS
+	zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+		&self->input, ZSTD_e_continue);
+	Py_END_ALLOW_THREADS
+
+	/* The input buffer currently points to memory managed by Python
+	(readBuffer). This object was allocated by this function. If it wasn't
+	fully consumed, we need to release it in a subsequent function call.
+	If it is fully consumed, do that now.
+	*/
+	if (self->input.pos == self->input.size) {
+		self->input.src = NULL;
+		self->input.pos = 0;
+		self->input.size = 0;
+		Py_XDECREF(self->readResult);
+		self->readResult = NULL;
+	}
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "zstd compress error: %s", ZSTD_getErrorName(zresult));
+		return NULL;
+	}
+
+	assert(self->input.pos <= self->input.size);
+
+	/* If we didn't write anything, start the process over. */
+	if (0 == self->output.pos) {
+		goto feedcompressor;
+	}
+
+	chunk = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+	self->output.pos = 0;
+	return chunk;
+}
+
+PyTypeObject ZstdCompressorIteratorType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdCompressorIterator",   /* tp_name */
+	sizeof(ZstdCompressorIterator),  /* tp_basicsize */
+	0,                               /* tp_itemsize */
+	(destructor)ZstdCompressorIterator_dealloc, /* tp_dealloc */
+	0,                               /* tp_print */
+	0,                               /* tp_getattr */
+	0,                               /* tp_setattr */
+	0,                               /* tp_compare */
+	0,                               /* tp_repr */
+	0,                               /* tp_as_number */
+	0,                               /* tp_as_sequence */
+	0,                               /* tp_as_mapping */
+	0,                               /* tp_hash */
+	0,                               /* tp_call */
+	0,                               /* tp_str */
+	0,                               /* tp_getattro */
+	0,                               /* tp_setattro */
+	0,                               /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdCompressorIterator__doc__,   /* tp_doc */
+	0,                               /* tp_traverse */
+	0,                               /* tp_clear */
+	0,                               /* tp_richcompare */
+	0,                               /* tp_weaklistoffset */
+	ZstdCompressorIterator_iter,     /* tp_iter */
+	(iternextfunc)ZstdCompressorIterator_iternext, /* tp_iternext */
+	0,                               /* tp_methods */
+	0,                               /* tp_members */
+	0,                               /* tp_getset */
+	0,                               /* tp_base */
+	0,                               /* tp_dict */
+	0,                               /* tp_descr_get */
+	0,                               /* tp_descr_set */
+	0,                               /* tp_dictoffset */
+	0,                               /* tp_init */
+	0,                               /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void compressoriterator_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdCompressorIteratorType) = &PyType_Type;
+	if (PyType_Ready(&ZstdCompressorIteratorType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/constants.c b/contrib/python/zstandard/py2/c-ext/constants.c
new file mode 100644
index 00000000000..bafdf1e469a
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/constants.c
@@ -0,0 +1,109 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+static char frame_header[] = {
+	'\x28',
+	'\xb5',
+	'\x2f',
+	'\xfd',
+};
+
+void constants_module_init(PyObject* mod) {
+	PyObject* version;
+	PyObject* zstdVersion;
+	PyObject* frameHeader;
+
+#if PY_MAJOR_VERSION >= 3
+	version = PyUnicode_FromString(PYTHON_ZSTANDARD_VERSION);
+#else
+	version = PyString_FromString(PYTHON_ZSTANDARD_VERSION);
+#endif
+	PyModule_AddObject(mod, "__version__", version);
+
+	ZstdError = PyErr_NewException("zstd.ZstdError", NULL, NULL);
+	PyModule_AddObject(mod, "ZstdError", ZstdError);
+
+	PyModule_AddIntConstant(mod, "FLUSH_BLOCK", 0);
+	PyModule_AddIntConstant(mod, "FLUSH_FRAME", 1);
+
+	PyModule_AddIntConstant(mod, "COMPRESSOBJ_FLUSH_FINISH", compressorobj_flush_finish);
+	PyModule_AddIntConstant(mod, "COMPRESSOBJ_FLUSH_BLOCK", compressorobj_flush_block);
+
+	/* For now, the version is a simple tuple instead of a dedicated type. */
+	zstdVersion = PyTuple_New(3);
+	PyTuple_SetItem(zstdVersion, 0, PyLong_FromLong(ZSTD_VERSION_MAJOR));
+	PyTuple_SetItem(zstdVersion, 1, PyLong_FromLong(ZSTD_VERSION_MINOR));
+	PyTuple_SetItem(zstdVersion, 2, PyLong_FromLong(ZSTD_VERSION_RELEASE));
+	PyModule_AddObject(mod, "ZSTD_VERSION", zstdVersion);
+
+	frameHeader = PyBytes_FromStringAndSize(frame_header, sizeof(frame_header));
+	if (frameHeader) {
+		PyModule_AddObject(mod, "FRAME_HEADER", frameHeader);
+	}
+	else {
+		PyErr_Format(PyExc_ValueError, "could not create frame header object");
+	}
+
+	PyModule_AddObject(mod, "CONTENTSIZE_UNKNOWN",
+		PyLong_FromUnsignedLongLong(ZSTD_CONTENTSIZE_UNKNOWN));
+	PyModule_AddObject(mod, "CONTENTSIZE_ERROR",
+		PyLong_FromUnsignedLongLong(ZSTD_CONTENTSIZE_ERROR));
+
+	PyModule_AddIntConstant(mod, "MAX_COMPRESSION_LEVEL", ZSTD_maxCLevel());
+	PyModule_AddIntConstant(mod, "COMPRESSION_RECOMMENDED_INPUT_SIZE",
+		(long)ZSTD_CStreamInSize());
+	PyModule_AddIntConstant(mod, "COMPRESSION_RECOMMENDED_OUTPUT_SIZE",
+		(long)ZSTD_CStreamOutSize());
+	PyModule_AddIntConstant(mod, "DECOMPRESSION_RECOMMENDED_INPUT_SIZE",
+		(long)ZSTD_DStreamInSize());
+	PyModule_AddIntConstant(mod, "DECOMPRESSION_RECOMMENDED_OUTPUT_SIZE",
+		(long)ZSTD_DStreamOutSize());
+
+	PyModule_AddIntConstant(mod, "MAGIC_NUMBER", ZSTD_MAGICNUMBER);
+	PyModule_AddIntConstant(mod, "BLOCKSIZELOG_MAX", ZSTD_BLOCKSIZELOG_MAX);
+	PyModule_AddIntConstant(mod, "BLOCKSIZE_MAX", ZSTD_BLOCKSIZE_MAX);
+	PyModule_AddIntConstant(mod, "WINDOWLOG_MIN", ZSTD_WINDOWLOG_MIN);
+	PyModule_AddIntConstant(mod, "WINDOWLOG_MAX", ZSTD_WINDOWLOG_MAX);
+	PyModule_AddIntConstant(mod, "CHAINLOG_MIN", ZSTD_CHAINLOG_MIN);
+	PyModule_AddIntConstant(mod, "CHAINLOG_MAX", ZSTD_CHAINLOG_MAX);
+	PyModule_AddIntConstant(mod, "HASHLOG_MIN", ZSTD_HASHLOG_MIN);
+	PyModule_AddIntConstant(mod, "HASHLOG_MAX", ZSTD_HASHLOG_MAX);
+	PyModule_AddIntConstant(mod, "SEARCHLOG_MIN", ZSTD_SEARCHLOG_MIN);
+	PyModule_AddIntConstant(mod, "SEARCHLOG_MAX", ZSTD_SEARCHLOG_MAX);
+	PyModule_AddIntConstant(mod, "MINMATCH_MIN", ZSTD_MINMATCH_MIN);
+	PyModule_AddIntConstant(mod, "MINMATCH_MAX", ZSTD_MINMATCH_MAX);
+	/* TODO SEARCHLENGTH_* is deprecated. */
+	PyModule_AddIntConstant(mod, "SEARCHLENGTH_MIN", ZSTD_MINMATCH_MIN);
+	PyModule_AddIntConstant(mod, "SEARCHLENGTH_MAX", ZSTD_MINMATCH_MAX);
+	PyModule_AddIntConstant(mod, "TARGETLENGTH_MIN", ZSTD_TARGETLENGTH_MIN);
+	PyModule_AddIntConstant(mod, "TARGETLENGTH_MAX", ZSTD_TARGETLENGTH_MAX);
+	PyModule_AddIntConstant(mod, "LDM_MINMATCH_MIN", ZSTD_LDM_MINMATCH_MIN);
+	PyModule_AddIntConstant(mod, "LDM_MINMATCH_MAX", ZSTD_LDM_MINMATCH_MAX);
+	PyModule_AddIntConstant(mod, "LDM_BUCKETSIZELOG_MAX", ZSTD_LDM_BUCKETSIZELOG_MAX);
+
+	PyModule_AddIntConstant(mod, "STRATEGY_FAST", ZSTD_fast);
+	PyModule_AddIntConstant(mod, "STRATEGY_DFAST", ZSTD_dfast);
+	PyModule_AddIntConstant(mod, "STRATEGY_GREEDY", ZSTD_greedy);
+	PyModule_AddIntConstant(mod, "STRATEGY_LAZY", ZSTD_lazy);
+	PyModule_AddIntConstant(mod, "STRATEGY_LAZY2", ZSTD_lazy2);
+	PyModule_AddIntConstant(mod, "STRATEGY_BTLAZY2", ZSTD_btlazy2);
+	PyModule_AddIntConstant(mod, "STRATEGY_BTOPT", ZSTD_btopt);
+	PyModule_AddIntConstant(mod, "STRATEGY_BTULTRA", ZSTD_btultra);
+	PyModule_AddIntConstant(mod, "STRATEGY_BTULTRA2", ZSTD_btultra2);
+
+	PyModule_AddIntConstant(mod, "DICT_TYPE_AUTO", ZSTD_dct_auto);
+	PyModule_AddIntConstant(mod, "DICT_TYPE_RAWCONTENT", ZSTD_dct_rawContent);
+	PyModule_AddIntConstant(mod, "DICT_TYPE_FULLDICT", ZSTD_dct_fullDict);
+
+	PyModule_AddIntConstant(mod, "FORMAT_ZSTD1", ZSTD_f_zstd1);
+	PyModule_AddIntConstant(mod, "FORMAT_ZSTD1_MAGICLESS", ZSTD_f_zstd1_magicless);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/decompressionreader.c b/contrib/python/zstandard/py2/c-ext/decompressionreader.c
new file mode 100644
index 00000000000..792852f2abe
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/decompressionreader.c
@@ -0,0 +1,781 @@
+/**
+* Copyright (c) 2017-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+static void set_unsupported_operation(void) {
+	PyObject* iomod;
+	PyObject* exc;
+
+	iomod = PyImport_ImportModule("io");
+	if (NULL == iomod) {
+		return;
+	}
+
+	exc = PyObject_GetAttrString(iomod, "UnsupportedOperation");
+	if (NULL == exc) {
+		Py_DECREF(iomod);
+		return;
+	}
+
+	PyErr_SetNone(exc);
+	Py_DECREF(exc);
+	Py_DECREF(iomod);
+}
+
+static void reader_dealloc(ZstdDecompressionReader* self) {
+	Py_XDECREF(self->decompressor);
+	Py_XDECREF(self->reader);
+
+	if (self->buffer.buf) {
+		PyBuffer_Release(&self->buffer);
+	}
+
+	PyObject_Del(self);
+}
+
+static ZstdDecompressionReader* reader_enter(ZstdDecompressionReader* self) {
+	if (self->entered) {
+		PyErr_SetString(PyExc_ValueError, "cannot __enter__ multiple times");
+		return NULL;
+	}
+
+	self->entered = 1;
+
+	Py_INCREF(self);
+	return self;
+}
+
+static PyObject* reader_exit(ZstdDecompressionReader* self, PyObject* args) {
+	PyObject* exc_type;
+	PyObject* exc_value;
+	PyObject* exc_tb;
+
+	if (!PyArg_ParseTuple(args, "OOO:__exit__", &exc_type, &exc_value, &exc_tb)) {
+		return NULL;
+	}
+
+	self->entered = 0;
+	self->closed = 1;
+
+	/* Release resources. */
+	Py_CLEAR(self->reader);
+	if (self->buffer.buf) {
+		PyBuffer_Release(&self->buffer);
+		memset(&self->buffer, 0, sizeof(self->buffer));
+	}
+
+	Py_CLEAR(self->decompressor);
+
+	Py_RETURN_FALSE;
+}
+
+static PyObject* reader_readable(PyObject* self) {
+	Py_RETURN_TRUE;
+}
+
+static PyObject* reader_writable(PyObject* self) {
+	Py_RETURN_FALSE;
+}
+
+static PyObject* reader_seekable(PyObject* self) {
+	Py_RETURN_TRUE;
+}
+
+static PyObject* reader_close(ZstdDecompressionReader* self) {
+	self->closed = 1;
+	Py_RETURN_NONE;
+}
+
+static PyObject* reader_flush(PyObject* self) {
+	Py_RETURN_NONE;
+}
+
+static PyObject* reader_isatty(PyObject* self) {
+	Py_RETURN_FALSE;
+}
+
+/**
+ * Read available input.
+ *
+ * Returns 0 if no data was added to input.
+ * Returns 1 if new input data is available.
+ * Returns -1 on error and sets a Python exception as a side-effect.
+ */
+int read_decompressor_input(ZstdDecompressionReader* self) {
+	if (self->finishedInput) {
+		return 0;
+	}
+
+	if (self->input.pos != self->input.size) {
+		return 0;
+	}
+
+	if (self->reader) {
+        Py_buffer buffer;
+
+        assert(self->readResult == NULL);
+        self->readResult = PyObject_CallMethod(self->reader, "read",
+            "k", self->readSize);
+        if (NULL == self->readResult) {
+            return -1;
+        }
+
+        memset(&buffer, 0, sizeof(buffer));
+
+        if (0 != PyObject_GetBuffer(self->readResult, &buffer, PyBUF_CONTIG_RO)) {
+            return -1;
+        }
+
+        /* EOF */
+        if (0 == buffer.len) {
+            self->finishedInput = 1;
+            Py_CLEAR(self->readResult);
+        }
+        else {
+            self->input.src = buffer.buf;
+            self->input.size = buffer.len;
+            self->input.pos = 0;
+        }
+
+        PyBuffer_Release(&buffer);
+	}
+	else {
+		assert(self->buffer.buf);
+        /*
+         * We should only get here once since expectation is we always
+         * exhaust input buffer before reading again.
+         */
+        assert(self->input.src == NULL);
+
+		self->input.src = self->buffer.buf;
+        self->input.size = self->buffer.len;
+        self->input.pos = 0;
+	}
+
+	return 1;
+}
+
+/**
+ * Decompresses available input into an output buffer.
+ *
+ * Returns 0 if we need more input.
+ * Returns 1 if output buffer should be emitted.
+ * Returns -1 on error and sets a Python exception.
+ */
+int decompress_input(ZstdDecompressionReader* self, ZSTD_outBuffer* output) {
+	size_t zresult;
+
+	if (self->input.pos >= self->input.size) {
+		return 0;
+	}
+
+	Py_BEGIN_ALLOW_THREADS
+	zresult = ZSTD_decompressStream(self->decompressor->dctx, output, &self->input);
+	Py_END_ALLOW_THREADS
+
+	/* Input exhausted. Clear our state tracking. */
+	if (self->input.pos == self->input.size) {
+		memset(&self->input, 0, sizeof(self->input));
+		Py_CLEAR(self->readResult);
+
+		if (self->buffer.buf) {
+			self->finishedInput = 1;
+		}
+	}
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "zstd decompress error: %s", ZSTD_getErrorName(zresult));
+		return -1;
+	}
+
+	/* We fulfilled the full read request. Signal to emit. */
+	if (output->pos && output->pos == output->size) {
+		return 1;
+	}
+	/* We're at the end of a frame and we aren't allowed to return data
+	   spanning frames. */
+	else if (output->pos && zresult == 0 && !self->readAcrossFrames) {
+		return 1;
+	}
+
+	/* There is more room in the output. Signal to collect more data. */
+	return 0;
+}
+
+static PyObject* reader_read(ZstdDecompressionReader* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"size",
+		NULL
+	};
+
+	Py_ssize_t size = -1;
+	PyObject* result = NULL;
+	char* resultBuffer;
+	Py_ssize_t resultSize;
+	ZSTD_outBuffer output;
+	int decompressResult, readResult;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n", kwlist, &size)) {
+		return NULL;
+	}
+
+	if (size < -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot read negative amounts less than -1");
+		return NULL;
+	}
+
+	if (size == -1) {
+		return PyObject_CallMethod((PyObject*)self, "readall", NULL);
+	}
+
+	if (self->finishedOutput || size == 0) {
+		return PyBytes_FromStringAndSize("", 0);
+	}
+
+	result = PyBytes_FromStringAndSize(NULL, size);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+	output.dst = resultBuffer;
+	output.size = resultSize;
+	output.pos = 0;
+
+readinput:
+
+	decompressResult = decompress_input(self, &output);
+
+	if (-1 == decompressResult) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+	else if (0 == decompressResult) { }
+	else if (1 == decompressResult) {
+		self->bytesDecompressed += output.pos;
+
+		if (output.pos != output.size) {
+			if (safe_pybytes_resize(&result, output.pos)) {
+				Py_XDECREF(result);
+				return NULL;
+			}
+		}
+		return result;
+	}
+	else {
+		assert(0);
+	}
+
+	readResult = read_decompressor_input(self);
+
+	if (-1 == readResult) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+	else if (0 == readResult) {}
+	else if (1 == readResult) {}
+	else {
+		assert(0);
+	}
+
+	if (self->input.size) {
+		goto readinput;
+	}
+
+	/* EOF */
+	self->bytesDecompressed += output.pos;
+
+	if (safe_pybytes_resize(&result, output.pos)) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+
+	return result;
+}
+
+static PyObject* reader_read1(ZstdDecompressionReader* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"size",
+		NULL
+	};
+
+	Py_ssize_t size = -1;
+	PyObject* result = NULL;
+	char* resultBuffer;
+	Py_ssize_t resultSize;
+	ZSTD_outBuffer output;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n", kwlist, &size)) {
+		return NULL;
+	}
+
+	if (size < -1) {
+		PyErr_SetString(PyExc_ValueError, "cannot read negative amounts less than -1");
+		return NULL;
+	}
+
+	if (self->finishedOutput || size == 0) {
+		return PyBytes_FromStringAndSize("", 0);
+	}
+
+	if (size == -1) {
+		size = ZSTD_DStreamOutSize();
+	}
+
+	result = PyBytes_FromStringAndSize(NULL, size);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+	output.dst = resultBuffer;
+	output.size = resultSize;
+	output.pos = 0;
+
+	/* read1() is supposed to use at most 1 read() from the underlying stream.
+	 * However, we can't satisfy this requirement with decompression due to the
+	 * nature of how decompression works. Our strategy is to read + decompress
+	 * until we get any output, at which point we return. This satisfies the
+	 * intent of the read1() API to limit read operations.
+	 */
+	while (!self->finishedInput) {
+		int readResult, decompressResult;
+
+		readResult = read_decompressor_input(self);
+		if (-1 == readResult) {
+			Py_XDECREF(result);
+			return NULL;
+		}
+		else if (0 == readResult || 1 == readResult) { }
+		else {
+			assert(0);
+		}
+
+		decompressResult = decompress_input(self, &output);
+
+		if (-1 == decompressResult) {
+			Py_XDECREF(result);
+			return NULL;
+		}
+		else if (0 == decompressResult || 1 == decompressResult) { }
+		else {
+			assert(0);
+		}
+
+		if (output.pos) {
+		    break;
+		}
+	}
+
+	self->bytesDecompressed += output.pos;
+	if (safe_pybytes_resize(&result, output.pos)) {
+		Py_XDECREF(result);
+		return NULL;
+	}
+
+	return result;
+}
+
+static PyObject* reader_readinto(ZstdDecompressionReader* self, PyObject* args) {
+	Py_buffer dest;
+	ZSTD_outBuffer output;
+	int decompressResult, readResult;
+	PyObject* result = NULL;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (self->finishedOutput) {
+		return PyLong_FromLong(0);
+	}
+
+	if (!PyArg_ParseTuple(args, "w*:readinto", &dest)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&dest, 'C') || dest.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"destination buffer should be contiguous and have at most one dimension");
+	    goto finally;
+	}
+
+	output.dst = dest.buf;
+	output.size = dest.len;
+	output.pos = 0;
+
+readinput:
+
+	decompressResult = decompress_input(self, &output);
+
+	if (-1 == decompressResult) {
+		goto finally;
+	}
+	else if (0 == decompressResult) { }
+	else if (1 == decompressResult) {
+		self->bytesDecompressed += output.pos;
+		result = PyLong_FromSize_t(output.pos);
+		goto finally;
+	}
+	else {
+		assert(0);
+	}
+
+	readResult = read_decompressor_input(self);
+
+	if (-1 == readResult) {
+		goto finally;
+	}
+	else if (0 == readResult) {}
+	else if (1 == readResult) {}
+	else {
+		assert(0);
+	}
+
+	if (self->input.size) {
+		goto readinput;
+	}
+
+	/* EOF */
+	self->bytesDecompressed += output.pos;
+	result = PyLong_FromSize_t(output.pos);
+
+finally:
+	PyBuffer_Release(&dest);
+
+	return result;
+}
+
+static PyObject* reader_readinto1(ZstdDecompressionReader* self, PyObject* args) {
+	Py_buffer dest;
+	ZSTD_outBuffer output;
+	PyObject* result = NULL;
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (self->finishedOutput) {
+		return PyLong_FromLong(0);
+	}
+
+	if (!PyArg_ParseTuple(args, "w*:readinto1", &dest)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&dest, 'C') || dest.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"destination buffer should be contiguous and have at most one dimension");
+	    goto finally;
+	}
+
+	output.dst = dest.buf;
+	output.size = dest.len;
+	output.pos = 0;
+
+	while (!self->finishedInput && !self->finishedOutput) {
+		int decompressResult, readResult;
+
+		readResult = read_decompressor_input(self);
+
+		if (-1 == readResult) {
+			goto finally;
+		}
+		else if (0 == readResult || 1 == readResult) {}
+		else {
+			assert(0);
+		}
+
+		decompressResult = decompress_input(self, &output);
+
+		if (-1 == decompressResult) {
+			goto finally;
+		}
+		else if (0 == decompressResult || 1 == decompressResult) {}
+		else {
+			assert(0);
+		}
+
+		if (output.pos) {
+			break;
+		}
+	}
+
+	self->bytesDecompressed += output.pos;
+	result = PyLong_FromSize_t(output.pos);
+
+finally:
+	PyBuffer_Release(&dest);
+
+	return result;
+}
+
+static PyObject* reader_readall(PyObject* self) {
+	PyObject* chunks = NULL;
+	PyObject* empty = NULL;
+	PyObject* result = NULL;
+
+	/* Our strategy is to collect chunks into a list then join all the
+	 * chunks at the end. We could potentially use e.g. an io.BytesIO. But
+	 * this feels simple enough to implement and avoids potentially expensive
+	 * reallocations of large buffers.
+	 */
+	chunks = PyList_New(0);
+	if (NULL == chunks) {
+		return NULL;
+	}
+
+	while (1) {
+		PyObject* chunk = PyObject_CallMethod(self, "read", "i", 1048576);
+		if (NULL == chunk) {
+			Py_DECREF(chunks);
+			return NULL;
+		}
+
+		if (!PyBytes_Size(chunk)) {
+			Py_DECREF(chunk);
+			break;
+		}
+
+		if (PyList_Append(chunks, chunk)) {
+			Py_DECREF(chunk);
+			Py_DECREF(chunks);
+			return NULL;
+		}
+
+		Py_DECREF(chunk);
+	}
+
+	empty = PyBytes_FromStringAndSize("", 0);
+	if (NULL == empty) {
+		Py_DECREF(chunks);
+		return NULL;
+	}
+
+	result = PyObject_CallMethod(empty, "join", "O", chunks);
+
+	Py_DECREF(empty);
+	Py_DECREF(chunks);
+
+	return result;
+}
+
+static PyObject* reader_readline(PyObject* self) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_readlines(PyObject* self) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_seek(ZstdDecompressionReader* self, PyObject* args) {
+	Py_ssize_t pos;
+	int whence = 0;
+	unsigned long long readAmount = 0;
+	size_t defaultOutSize = ZSTD_DStreamOutSize();
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (!PyArg_ParseTuple(args, "n|i:seek", &pos, &whence)) {
+		return NULL;
+	}
+
+	if (whence == SEEK_SET) {
+		if (pos < 0) {
+			PyErr_SetString(PyExc_ValueError,
+				"cannot seek to negative position with SEEK_SET");
+			return NULL;
+		}
+
+		if ((unsigned long long)pos < self->bytesDecompressed) {
+			PyErr_SetString(PyExc_ValueError,
+				"cannot seek zstd decompression stream backwards");
+			return NULL;
+		}
+
+		readAmount = pos - self->bytesDecompressed;
+	}
+	else if (whence == SEEK_CUR) {
+		if (pos < 0) {
+			PyErr_SetString(PyExc_ValueError,
+				"cannot seek zstd decompression stream backwards");
+			return NULL;
+		}
+
+		readAmount = pos;
+	}
+	else if (whence == SEEK_END) {
+		/* We /could/ support this with pos==0. But let's not do that until someone
+		   needs it. */
+		PyErr_SetString(PyExc_ValueError,
+			"zstd decompression streams cannot be seeked with SEEK_END");
+		return NULL;
+	}
+
+	/* It is a bit inefficient to do this via the Python API. But since there
+	   is a bit of state tracking involved to read from this type, it is the
+	   easiest to implement. */
+	while (readAmount) {
+		Py_ssize_t readSize;
+		PyObject* readResult = PyObject_CallMethod((PyObject*)self, "read", "K",
+			readAmount < defaultOutSize ? readAmount : defaultOutSize);
+
+		if (!readResult) {
+			return NULL;
+		}
+
+		readSize = PyBytes_GET_SIZE(readResult);
+
+		Py_CLEAR(readResult);
+
+		/* Empty read means EOF. */
+		if (!readSize) {
+			break;
+		}
+
+		readAmount -= readSize;
+	}
+
+	return PyLong_FromUnsignedLongLong(self->bytesDecompressed);
+}
+
+static PyObject* reader_tell(ZstdDecompressionReader* self) {
+	/* TODO should this raise OSError since stream isn't seekable? */
+	return PyLong_FromUnsignedLongLong(self->bytesDecompressed);
+}
+
+static PyObject* reader_write(PyObject* self, PyObject* args) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_writelines(PyObject* self, PyObject* args) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_iter(PyObject* self) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyObject* reader_iternext(PyObject* self) {
+	set_unsupported_operation();
+	return NULL;
+}
+
+static PyMethodDef reader_methods[] = {
+	{ "__enter__", (PyCFunction)reader_enter, METH_NOARGS,
+	PyDoc_STR("Enter a compression context") },
+	{ "__exit__", (PyCFunction)reader_exit, METH_VARARGS,
+	PyDoc_STR("Exit a compression context") },
+	{ "close", (PyCFunction)reader_close, METH_NOARGS,
+	PyDoc_STR("Close the stream so it cannot perform any more operations") },
+	{ "flush", (PyCFunction)reader_flush, METH_NOARGS, PyDoc_STR("no-ops") },
+	{ "isatty", (PyCFunction)reader_isatty, METH_NOARGS, PyDoc_STR("Returns False") },
+	{ "readable", (PyCFunction)reader_readable, METH_NOARGS,
+	PyDoc_STR("Returns True") },
+	{ "read", (PyCFunction)reader_read, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("read compressed data") },
+	{ "read1", (PyCFunction)reader_read1, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("read compressed data") },
+	{ "readinto", (PyCFunction)reader_readinto, METH_VARARGS, NULL },
+	{ "readinto1", (PyCFunction)reader_readinto1, METH_VARARGS, NULL },
+	{ "readall", (PyCFunction)reader_readall, METH_NOARGS, PyDoc_STR("Not implemented") },
+	{ "readline", (PyCFunction)reader_readline, METH_NOARGS, PyDoc_STR("Not implemented") },
+	{ "readlines", (PyCFunction)reader_readlines, METH_NOARGS, PyDoc_STR("Not implemented") },
+	{ "seek", (PyCFunction)reader_seek, METH_VARARGS, PyDoc_STR("Seek the stream") },
+	{ "seekable", (PyCFunction)reader_seekable, METH_NOARGS,
+	PyDoc_STR("Returns True") },
+	{ "tell", (PyCFunction)reader_tell, METH_NOARGS,
+	PyDoc_STR("Returns current number of bytes compressed") },
+	{ "writable", (PyCFunction)reader_writable, METH_NOARGS,
+	PyDoc_STR("Returns False") },
+	{ "write", (PyCFunction)reader_write, METH_VARARGS, PyDoc_STR("unsupported operation") },
+	{ "writelines", (PyCFunction)reader_writelines, METH_VARARGS, PyDoc_STR("unsupported operation") },
+	{ NULL, NULL }
+};
+
+static PyMemberDef reader_members[] = {
+	{ "closed", T_BOOL, offsetof(ZstdDecompressionReader, closed),
+	  READONLY, "whether stream is closed" },
+	{ NULL }
+};
+
+PyTypeObject ZstdDecompressionReaderType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdDecompressionReader", /* tp_name */
+	sizeof(ZstdDecompressionReader), /* tp_basicsize */
+	0, /* tp_itemsize */
+	(destructor)reader_dealloc, /* tp_dealloc */
+	0, /* tp_print */
+	0, /* tp_getattr */
+	0, /* tp_setattr */
+	0, /* tp_compare */
+	0, /* tp_repr */
+	0, /* tp_as_number */
+	0, /* tp_as_sequence */
+	0, /* tp_as_mapping */
+	0, /* tp_hash */
+	0, /* tp_call */
+	0, /* tp_str */
+	0, /* tp_getattro */
+	0, /* tp_setattro */
+	0, /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT, /* tp_flags */
+	0, /* tp_doc */
+	0, /* tp_traverse */
+	0, /* tp_clear */
+	0, /* tp_richcompare */
+	0, /* tp_weaklistoffset */
+	reader_iter, /* tp_iter */
+	reader_iternext, /* tp_iternext */
+	reader_methods, /* tp_methods */
+	reader_members, /* tp_members */
+	0, /* tp_getset */
+	0, /* tp_base */
+	0, /* tp_dict */
+	0, /* tp_descr_get */
+	0, /* tp_descr_set */
+	0, /* tp_dictoffset */
+	0, /* tp_init */
+	0, /* tp_alloc */
+	PyType_GenericNew, /* tp_new */
+};
+
+
+void decompressionreader_module_init(PyObject* mod) {
+	/* TODO make reader a sub-class of io.RawIOBase */
+
+	Py_TYPE(&ZstdDecompressionReaderType) = &PyType_Type;
+	if (PyType_Ready(&ZstdDecompressionReaderType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/decompressionwriter.c b/contrib/python/zstandard/py2/c-ext/decompressionwriter.c
new file mode 100644
index 00000000000..7df750ab06e
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/decompressionwriter.c
@@ -0,0 +1,295 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(ZstdDecompressionWriter__doc,
+"""A context manager used for writing decompressed output.\n"
+);
+
+static void ZstdDecompressionWriter_dealloc(ZstdDecompressionWriter* self) {
+	Py_XDECREF(self->decompressor);
+	Py_XDECREF(self->writer);
+
+	PyObject_Del(self);
+}
+
+static PyObject* ZstdDecompressionWriter_enter(ZstdDecompressionWriter* self) {
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (self->entered) {
+		PyErr_SetString(ZstdError, "cannot __enter__ multiple times");
+		return NULL;
+	}
+
+	self->entered = 1;
+
+	Py_INCREF(self);
+	return (PyObject*)self;
+}
+
+static PyObject* ZstdDecompressionWriter_exit(ZstdDecompressionWriter* self, PyObject* args) {
+	self->entered = 0;
+
+	if (NULL == PyObject_CallMethod((PyObject*)self, "close", NULL)) {
+		return NULL;
+	}
+
+	Py_RETURN_FALSE;
+}
+
+static PyObject* ZstdDecompressionWriter_memory_size(ZstdDecompressionWriter* self) {
+	return PyLong_FromSize_t(ZSTD_sizeof_DCtx(self->decompressor->dctx));
+}
+
+static PyObject* ZstdDecompressionWriter_write(ZstdDecompressionWriter* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	PyObject* result = NULL;
+	Py_buffer source;
+	size_t zresult = 0;
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	PyObject* res;
+	Py_ssize_t totalWrite = 0;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:write",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:write",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	output.dst = PyMem_Malloc(self->outSize);
+	if (!output.dst) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+	output.size = self->outSize;
+	output.pos = 0;
+
+	input.src = source.buf;
+	input.size = source.len;
+	input.pos = 0;
+
+	while (input.pos < (size_t)source.len) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_decompressStream(self->decompressor->dctx, &output, &input);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyMem_Free(output.dst);
+			PyErr_Format(ZstdError, "zstd decompress error: %s",
+				ZSTD_getErrorName(zresult));
+			goto finally;
+		}
+
+		if (output.pos) {
+#if PY_MAJOR_VERSION >= 3
+			res = PyObject_CallMethod(self->writer, "write", "y#",
+#else
+			res = PyObject_CallMethod(self->writer, "write", "s#",
+#endif
+				output.dst, output.pos);
+			Py_XDECREF(res);
+			totalWrite += output.pos;
+			output.pos = 0;
+		}
+	}
+
+	PyMem_Free(output.dst);
+
+	if (self->writeReturnRead) {
+		result = PyLong_FromSize_t(input.pos);
+	}
+	else {
+		result = PyLong_FromSsize_t(totalWrite);
+	}
+
+finally:
+	PyBuffer_Release(&source);
+	return result;
+}
+
+static PyObject* ZstdDecompressionWriter_close(ZstdDecompressionWriter* self) {
+	PyObject* result;
+
+	if (self->closed) {
+		Py_RETURN_NONE;
+	}
+
+	result = PyObject_CallMethod((PyObject*)self, "flush", NULL);
+	self->closed = 1;
+
+	if (NULL == result) {
+		return NULL;
+	}
+
+	/* Call close on underlying stream as well. */
+	if (PyObject_HasAttrString(self->writer, "close")) {
+		return PyObject_CallMethod(self->writer, "close", NULL);
+	}
+
+	Py_RETURN_NONE;
+}
+
+static PyObject* ZstdDecompressionWriter_fileno(ZstdDecompressionWriter* self) {
+	if (PyObject_HasAttrString(self->writer, "fileno")) {
+		return PyObject_CallMethod(self->writer, "fileno", NULL);
+	}
+	else {
+		PyErr_SetString(PyExc_OSError, "fileno not available on underlying writer");
+		return NULL;
+	}
+}
+
+static PyObject* ZstdDecompressionWriter_flush(ZstdDecompressionWriter* self) {
+	if (self->closed) {
+		PyErr_SetString(PyExc_ValueError, "stream is closed");
+		return NULL;
+	}
+
+	if (PyObject_HasAttrString(self->writer, "flush")) {
+		return PyObject_CallMethod(self->writer, "flush", NULL);
+	}
+	else {
+		Py_RETURN_NONE;
+	}
+}
+
+static PyObject* ZstdDecompressionWriter_false(PyObject* self, PyObject* args) {
+	Py_RETURN_FALSE;
+}
+
+static PyObject* ZstdDecompressionWriter_true(PyObject* self, PyObject* args) {
+	Py_RETURN_TRUE;
+}
+
+static PyObject* ZstdDecompressionWriter_unsupported(PyObject* self, PyObject* args, PyObject* kwargs) {
+	PyObject* iomod;
+	PyObject* exc;
+
+	iomod = PyImport_ImportModule("io");
+	if (NULL == iomod) {
+		return NULL;
+	}
+
+	exc = PyObject_GetAttrString(iomod, "UnsupportedOperation");
+	if (NULL == exc) {
+		Py_DECREF(iomod);
+		return NULL;
+	}
+
+	PyErr_SetNone(exc);
+	Py_DECREF(exc);
+	Py_DECREF(iomod);
+
+	return NULL;
+}
+
+static PyMethodDef ZstdDecompressionWriter_methods[] = {
+	{ "__enter__", (PyCFunction)ZstdDecompressionWriter_enter, METH_NOARGS,
+	PyDoc_STR("Enter a decompression context.") },
+	{ "__exit__", (PyCFunction)ZstdDecompressionWriter_exit, METH_VARARGS,
+	PyDoc_STR("Exit a decompression context.") },
+	{ "memory_size", (PyCFunction)ZstdDecompressionWriter_memory_size, METH_NOARGS,
+	PyDoc_STR("Obtain the memory size in bytes of the underlying decompressor.") },
+	{ "close", (PyCFunction)ZstdDecompressionWriter_close, METH_NOARGS, NULL },
+	{ "fileno", (PyCFunction)ZstdDecompressionWriter_fileno, METH_NOARGS, NULL },
+	{ "flush", (PyCFunction)ZstdDecompressionWriter_flush, METH_NOARGS, NULL },
+	{ "isatty", ZstdDecompressionWriter_false, METH_NOARGS, NULL },
+	{ "readable", ZstdDecompressionWriter_false, METH_NOARGS, NULL },
+	{ "readline", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readlines", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "seek", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "seekable", ZstdDecompressionWriter_false, METH_NOARGS, NULL },
+	{ "tell", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "truncate", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "writable", ZstdDecompressionWriter_true, METH_NOARGS, NULL },
+	{ "writelines" , (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "read", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readall", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "readinto", (PyCFunction)ZstdDecompressionWriter_unsupported, METH_VARARGS | METH_KEYWORDS, NULL },
+	{ "write", (PyCFunction)ZstdDecompressionWriter_write, METH_VARARGS | METH_KEYWORDS,
+	PyDoc_STR("Compress data") },
+	{ NULL, NULL }
+};
+
+static PyMemberDef ZstdDecompressionWriter_members[] = {
+	{ "closed", T_BOOL, offsetof(ZstdDecompressionWriter, closed), READONLY, NULL },
+	{ NULL }
+};
+
+PyTypeObject ZstdDecompressionWriterType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdDecompressionWriter", /* tp_name */
+	sizeof(ZstdDecompressionWriter),/* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)ZstdDecompressionWriter_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	0,                              /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdDecompressionWriter__doc,   /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	ZstdDecompressionWriter_methods,/* tp_methods */
+	ZstdDecompressionWriter_members,/* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	0,                              /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void decompressionwriter_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdDecompressionWriterType) = &PyType_Type;
+	if (PyType_Ready(&ZstdDecompressionWriterType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/decompressobj.c b/contrib/python/zstandard/py2/c-ext/decompressobj.c
new file mode 100644
index 00000000000..2a55e61f18b
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/decompressobj.c
@@ -0,0 +1,202 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(DecompressionObj__doc__,
+"Perform decompression using a standard library compatible API.\n"
+);
+
+static void DecompressionObj_dealloc(ZstdDecompressionObj* self) {
+	Py_XDECREF(self->decompressor);
+
+	PyObject_Del(self);
+}
+
+static PyObject* DecompressionObj_decompress(ZstdDecompressionObj* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	Py_buffer source;
+	size_t zresult;
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	PyObject* result = NULL;
+	Py_ssize_t resultSize = 0;
+
+	output.dst = NULL;
+
+	if (self->finished) {
+		PyErr_SetString(ZstdError, "cannot use a decompressobj multiple times");
+		return NULL;
+	}
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:decompress",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:decompress",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	/* Special case of empty input. Output will always be empty. */
+	if (source.len == 0) {
+		result = PyBytes_FromString("");
+		goto finally;
+	}
+
+	input.src = source.buf;
+	input.size = source.len;
+	input.pos = 0;
+
+	output.dst = PyMem_Malloc(self->outSize);
+	if (!output.dst) {
+		PyErr_NoMemory();
+		goto except;
+	}
+	output.size = self->outSize;
+	output.pos = 0;
+
+	while (1) {
+		Py_BEGIN_ALLOW_THREADS
+		zresult = ZSTD_decompressStream(self->decompressor->dctx, &output, &input);
+		Py_END_ALLOW_THREADS
+
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "zstd decompressor error: %s",
+				ZSTD_getErrorName(zresult));
+			goto except;
+		}
+
+		if (0 == zresult) {
+			self->finished = 1;
+		}
+
+		if (output.pos) {
+			if (result) {
+				resultSize = PyBytes_GET_SIZE(result);
+				if (-1 == safe_pybytes_resize(&result, resultSize + output.pos)) {
+					Py_XDECREF(result);
+					goto except;
+				}
+
+				memcpy(PyBytes_AS_STRING(result) + resultSize,
+					output.dst, output.pos);
+			}
+			else {
+				result = PyBytes_FromStringAndSize(output.dst, output.pos);
+				if (!result) {
+					goto except;
+				}
+			}
+		}
+
+		if (zresult == 0 || (input.pos == input.size && output.pos == 0)) {
+			break;
+		}
+
+		output.pos = 0;
+	}
+
+	if (!result) {
+		result = PyBytes_FromString("");
+	}
+
+	goto finally;
+
+except:
+	Py_CLEAR(result);
+
+finally:
+	PyMem_Free(output.dst);
+	PyBuffer_Release(&source);
+
+	return result;
+}
+
+static PyObject* DecompressionObj_flush(ZstdDecompressionObj* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"length",
+		NULL
+	};
+
+	PyObject* length = NULL;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|O:flush", kwlist, &length)) {
+	return NULL;
+	}
+
+	Py_RETURN_NONE;
+}
+
+static PyMethodDef DecompressionObj_methods[] = {
+	{ "decompress", (PyCFunction)DecompressionObj_decompress,
+	  METH_VARARGS | METH_KEYWORDS, PyDoc_STR("decompress data") },
+	{ "flush", (PyCFunction)DecompressionObj_flush,
+	  METH_VARARGS | METH_KEYWORDS, PyDoc_STR("no-op") },
+	{ NULL, NULL }
+};
+
+PyTypeObject ZstdDecompressionObjType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdDecompressionObj",    /* tp_name */
+	sizeof(ZstdDecompressionObj),   /* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)DecompressionObj_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	0,                              /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	DecompressionObj__doc__,        /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	DecompressionObj_methods,       /* tp_methods */
+	0,                              /* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	0,                              /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void decompressobj_module_init(PyObject* module) {
+	Py_TYPE(&ZstdDecompressionObjType) = &PyType_Type;
+	if (PyType_Ready(&ZstdDecompressionObjType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/decompressor.c b/contrib/python/zstandard/py2/c-ext/decompressor.c
new file mode 100644
index 00000000000..b1ef959ec58
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/decompressor.c
@@ -0,0 +1,1828 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+#include "pool.h"
+
+extern PyObject* ZstdError;
+
+/**
+ * Ensure the ZSTD_DCtx on a decompressor is initiated and ready for a new operation.
+ */
+int ensure_dctx(ZstdDecompressor* decompressor, int loadDict) {
+	size_t zresult;
+
+	ZSTD_DCtx_reset(decompressor->dctx, ZSTD_reset_session_only);
+
+	if (decompressor->maxWindowSize) {
+		zresult = ZSTD_DCtx_setMaxWindowSize(decompressor->dctx, decompressor->maxWindowSize);
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "unable to set max window size: %s",
+				ZSTD_getErrorName(zresult));
+			return 1;
+		}
+	}
+
+	zresult = ZSTD_DCtx_setFormat(decompressor->dctx, decompressor->format);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "unable to set decoding format: %s",
+			ZSTD_getErrorName(zresult));
+		return 1;
+	}
+
+	if (loadDict && decompressor->dict) {
+		if (ensure_ddict(decompressor->dict)) {
+			return 1;
+		}
+
+		zresult = ZSTD_DCtx_refDDict(decompressor->dctx, decompressor->dict->ddict);
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(ZstdError, "unable to reference prepared dictionary: %s",
+				ZSTD_getErrorName(zresult));
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+PyDoc_STRVAR(Decompressor__doc__,
+"ZstdDecompressor(dict_data=None)\n"
+"\n"
+"Create an object used to perform Zstandard decompression.\n"
+"\n"
+"An instance can perform multiple decompression operations."
+);
+
+static int Decompressor_init(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"dict_data",
+		"max_window_size",
+		"format",
+		NULL
+	};
+
+	ZstdCompressionDict* dict = NULL;
+	Py_ssize_t maxWindowSize = 0;
+	ZSTD_format_e format = ZSTD_f_zstd1;
+
+	self->dctx = NULL;
+	self->dict = NULL;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|O!nI:ZstdDecompressor", kwlist,
+		&ZstdCompressionDictType, &dict, &maxWindowSize, &format)) {
+		return -1;
+	}
+
+	self->dctx = ZSTD_createDCtx();
+	if (!self->dctx) {
+		PyErr_NoMemory();
+		goto except;
+	}
+
+	self->maxWindowSize = maxWindowSize;
+	self->format = format;
+
+	if (dict) {
+		self->dict = dict;
+		Py_INCREF(dict);
+	}
+
+	if (ensure_dctx(self, 1)) {
+		goto except;
+	}
+
+	return 0;
+
+except:
+	Py_CLEAR(self->dict);
+
+	if (self->dctx) {
+		ZSTD_freeDCtx(self->dctx);
+		self->dctx = NULL;
+	}
+
+	return -1;
+}
+
+static void Decompressor_dealloc(ZstdDecompressor* self) {
+	Py_CLEAR(self->dict);
+
+	if (self->dctx) {
+		ZSTD_freeDCtx(self->dctx);
+		self->dctx = NULL;
+	}
+
+	PyObject_Del(self);
+}
+
+PyDoc_STRVAR(Decompressor_memory_size__doc__,
+"memory_size() -- Size of decompression context, in bytes\n"
+);
+
+static PyObject* Decompressor_memory_size(ZstdDecompressor* self) {
+	if (self->dctx) {
+		return PyLong_FromSize_t(ZSTD_sizeof_DCtx(self->dctx));
+	}
+	else {
+		PyErr_SetString(ZstdError, "no decompressor context found; this should never happen");
+		return NULL;
+	}
+}
+
+PyDoc_STRVAR(Decompressor_copy_stream__doc__,
+	"copy_stream(ifh, ofh[, read_size=default, write_size=default]) -- decompress data between streams\n"
+	"\n"
+	"Compressed data will be read from ``ifh``, decompressed, and written to\n"
+	"``ofh``. ``ifh`` must have a ``read(size)`` method. ``ofh`` must have a\n"
+	"``write(data)`` method.\n"
+	"\n"
+	"The optional ``read_size`` and ``write_size`` arguments control the chunk\n"
+	"size of data that is ``read()`` and ``write()`` between streams. They default\n"
+	"to the default input and output sizes of zstd decompressor streams.\n"
+);
+
+static PyObject* Decompressor_copy_stream(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"ifh",
+		"ofh",
+		"read_size",
+		"write_size",
+		NULL
+	};
+
+	PyObject* source;
+	PyObject* dest;
+	size_t inSize = ZSTD_DStreamInSize();
+	size_t outSize = ZSTD_DStreamOutSize();
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	Py_ssize_t totalRead = 0;
+	Py_ssize_t totalWrite = 0;
+	char* readBuffer;
+	Py_ssize_t readSize;
+	PyObject* readResult = NULL;
+	PyObject* res = NULL;
+	size_t zresult = 0;
+	PyObject* writeResult;
+	PyObject* totalReadPy;
+	PyObject* totalWritePy;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "OO|kk:copy_stream", kwlist,
+		&source, &dest, &inSize, &outSize)) {
+		return NULL;
+	}
+
+	if (!PyObject_HasAttrString(source, "read")) {
+		PyErr_SetString(PyExc_ValueError, "first argument must have a read() method");
+		return NULL;
+	}
+
+	if (!PyObject_HasAttrString(dest, "write")) {
+		PyErr_SetString(PyExc_ValueError, "second argument must have a write() method");
+		return NULL;
+	}
+
+	/* Prevent free on uninitialized memory in finally. */
+	output.dst = NULL;
+
+	if (ensure_dctx(self, 1)) {
+		res = NULL;
+		goto finally;
+	}
+
+	output.dst = PyMem_Malloc(outSize);
+	if (!output.dst) {
+		PyErr_NoMemory();
+		res = NULL;
+		goto finally;
+	}
+	output.size = outSize;
+	output.pos = 0;
+
+	/* Read source stream until EOF */
+	while (1) {
+		readResult = PyObject_CallMethod(source, "read", "n", inSize);
+		if (!readResult) {
+			PyErr_SetString(ZstdError, "could not read() from source");
+			goto finally;
+		}
+
+		PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+
+		/* If no data was read, we're at EOF. */
+		if (0 == readSize) {
+			break;
+		}
+
+		totalRead += readSize;
+
+		/* Send data to decompressor */
+		input.src = readBuffer;
+		input.size = readSize;
+		input.pos = 0;
+
+		while (input.pos < input.size) {
+			Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_decompressStream(self->dctx, &output, &input);
+			Py_END_ALLOW_THREADS
+
+			if (ZSTD_isError(zresult)) {
+				PyErr_Format(ZstdError, "zstd decompressor error: %s",
+					ZSTD_getErrorName(zresult));
+				res = NULL;
+				goto finally;
+			}
+
+			if (output.pos) {
+#if PY_MAJOR_VERSION >= 3
+				writeResult = PyObject_CallMethod(dest, "write", "y#",
+#else
+				writeResult = PyObject_CallMethod(dest, "write", "s#",
+#endif
+					output.dst, output.pos);
+
+				Py_XDECREF(writeResult);
+				totalWrite += output.pos;
+				output.pos = 0;
+			}
+		}
+
+		Py_CLEAR(readResult);
+	}
+
+	/* Source stream is exhausted. Finish up. */
+
+	totalReadPy = PyLong_FromSsize_t(totalRead);
+	totalWritePy = PyLong_FromSsize_t(totalWrite);
+	res = PyTuple_Pack(2, totalReadPy, totalWritePy);
+	Py_DECREF(totalReadPy);
+	Py_DECREF(totalWritePy);
+
+finally:
+	if (output.dst) {
+		PyMem_Free(output.dst);
+	}
+
+	Py_XDECREF(readResult);
+
+	return res;
+}
+
+PyDoc_STRVAR(Decompressor_decompress__doc__,
+"decompress(data[, max_output_size=None]) -- Decompress data in its entirety\n"
+"\n"
+"This method will decompress the entirety of the argument and return the\n"
+"result.\n"
+"\n"
+"The input bytes are expected to contain a full Zstandard frame (something\n"
+"compressed with ``ZstdCompressor.compress()`` or similar). If the input does\n"
+"not contain a full frame, an exception will be raised.\n"
+"\n"
+"If the frame header of the compressed data does not contain the content size\n"
+"``max_output_size`` must be specified or ``ZstdError`` will be raised. An\n"
+"allocation of size ``max_output_size`` will be performed and an attempt will\n"
+"be made to perform decompression into that buffer. If the buffer is too\n"
+"small or cannot be allocated, ``ZstdError`` will be raised. The buffer will\n"
+"be resized if it is too large.\n"
+"\n"
+"Uncompressed data could be much larger than compressed data. As a result,\n"
+"calling this function could result in a very large memory allocation being\n"
+"performed to hold the uncompressed data. Therefore it is **highly**\n"
+"recommended to use a streaming decompression method instead of this one.\n"
+);
+
+PyObject* Decompressor_decompress(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		"max_output_size",
+		NULL
+	};
+
+	Py_buffer source;
+	Py_ssize_t maxOutputSize = 0;
+	unsigned long long decompressedSize;
+	size_t destCapacity;
+	PyObject* result = NULL;
+	size_t zresult;
+	ZSTD_outBuffer outBuffer;
+	ZSTD_inBuffer inBuffer;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|n:decompress",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*|n:decompress",
+#endif
+		kwlist, &source, &maxOutputSize)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	if (ensure_dctx(self, 1)) {
+		goto finally;
+	}
+
+	decompressedSize = ZSTD_getFrameContentSize(source.buf, source.len);
+
+	if (ZSTD_CONTENTSIZE_ERROR == decompressedSize) {
+		PyErr_SetString(ZstdError, "error determining content size from frame header");
+		goto finally;
+	}
+	/* Special case of empty frame. */
+	else if (0 == decompressedSize) {
+		result = PyBytes_FromStringAndSize("", 0);
+		goto finally;
+	}
+	/* Missing content size in frame header. */
+	if (ZSTD_CONTENTSIZE_UNKNOWN == decompressedSize) {
+		if (0 == maxOutputSize) {
+			PyErr_SetString(ZstdError, "could not determine content size in frame header");
+			goto finally;
+		}
+
+		result = PyBytes_FromStringAndSize(NULL, maxOutputSize);
+		destCapacity = maxOutputSize;
+		decompressedSize = 0;
+	}
+	/* Size is recorded in frame header. */
+	else {
+		assert(SIZE_MAX >= PY_SSIZE_T_MAX);
+		if (decompressedSize > PY_SSIZE_T_MAX) {
+			PyErr_SetString(ZstdError, "frame is too large to decompress on this platform");
+			goto finally;
+		}
+
+		result = PyBytes_FromStringAndSize(NULL, (Py_ssize_t)decompressedSize);
+		destCapacity = (size_t)decompressedSize;
+	}
+
+	if (!result) {
+		goto finally;
+	}
+
+	outBuffer.dst = PyBytes_AsString(result);
+	outBuffer.size = destCapacity;
+	outBuffer.pos = 0;
+
+	inBuffer.src = source.buf;
+	inBuffer.size = source.len;
+	inBuffer.pos = 0;
+
+	Py_BEGIN_ALLOW_THREADS
+	zresult = ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+	Py_END_ALLOW_THREADS
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "decompression error: %s", ZSTD_getErrorName(zresult));
+		Py_CLEAR(result);
+		goto finally;
+	}
+	else if (zresult) {
+		PyErr_Format(ZstdError, "decompression error: did not decompress full frame");
+		Py_CLEAR(result);
+		goto finally;
+	}
+	else if (decompressedSize && outBuffer.pos != decompressedSize) {
+		PyErr_Format(ZstdError, "decompression error: decompressed %zu bytes; expected %llu",
+			zresult, decompressedSize);
+		Py_CLEAR(result);
+		goto finally;
+	}
+	else if (outBuffer.pos < destCapacity) {
+		if (safe_pybytes_resize(&result, outBuffer.pos)) {
+			Py_CLEAR(result);
+			goto finally;
+		}
+	}
+
+finally:
+	PyBuffer_Release(&source);
+	return result;
+}
+
+PyDoc_STRVAR(Decompressor_decompressobj__doc__,
+"decompressobj([write_size=default])\n"
+"\n"
+"Incrementally feed data into a decompressor.\n"
+"\n"
+"The returned object exposes a ``decompress(data)`` method. This makes it\n"
+"compatible with ``zlib.decompressobj`` and ``bz2.BZ2Decompressor`` so that\n"
+"callers can swap in the zstd decompressor while using the same API.\n"
+);
+
+static ZstdDecompressionObj* Decompressor_decompressobj(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"write_size",
+		NULL
+	};
+
+	ZstdDecompressionObj* result = NULL;
+	size_t outSize = ZSTD_DStreamOutSize();
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|k:decompressobj", kwlist, &outSize)) {
+		return NULL;
+	}
+
+	if (!outSize) {
+		PyErr_SetString(PyExc_ValueError, "write_size must be positive");
+		return NULL;
+	}
+
+	result = (ZstdDecompressionObj*)PyObject_CallObject((PyObject*)&ZstdDecompressionObjType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	if (ensure_dctx(self, 1)) {
+		Py_DECREF(result);
+		return NULL;
+	}
+
+	result->decompressor = self;
+	Py_INCREF(result->decompressor);
+	result->outSize = outSize;
+
+	return result;
+}
+
+PyDoc_STRVAR(Decompressor_read_to_iter__doc__,
+"read_to_iter(reader[, read_size=default, write_size=default, skip_bytes=0])\n"
+"Read compressed data and return an iterator\n"
+"\n"
+"Returns an iterator of decompressed data chunks produced from reading from\n"
+"the ``reader``.\n"
+"\n"
+"Compressed data will be obtained from ``reader`` by calling the\n"
+"``read(size)`` method of it. The source data will be streamed into a\n"
+"decompressor. As decompressed data is available, it will be exposed to the\n"
+"returned iterator.\n"
+"\n"
+"Data is ``read()`` in chunks of size ``read_size`` and exposed to the\n"
+"iterator in chunks of size ``write_size``. The default values are the input\n"
+"and output sizes for a zstd streaming decompressor.\n"
+"\n"
+"There is also support for skipping the first ``skip_bytes`` of data from\n"
+"the source.\n"
+);
+
+static ZstdDecompressorIterator* Decompressor_read_to_iter(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"reader",
+		"read_size",
+		"write_size",
+		"skip_bytes",
+		NULL
+	};
+
+	PyObject* reader;
+	size_t inSize = ZSTD_DStreamInSize();
+	size_t outSize = ZSTD_DStreamOutSize();
+	ZstdDecompressorIterator* result;
+	size_t skipBytes = 0;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kkk:read_to_iter", kwlist,
+		&reader, &inSize, &outSize, &skipBytes)) {
+		return NULL;
+	}
+
+	if (skipBytes >= inSize) {
+		PyErr_SetString(PyExc_ValueError,
+			"skip_bytes must be smaller than read_size");
+		return NULL;
+	}
+
+	result = (ZstdDecompressorIterator*)PyObject_CallObject((PyObject*)&ZstdDecompressorIteratorType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	if (PyObject_HasAttrString(reader, "read")) {
+		result->reader = reader;
+		Py_INCREF(result->reader);
+	}
+	else if (1 == PyObject_CheckBuffer(reader)) {
+		/* Object claims it is a buffer. Try to get a handle to it. */
+		if (0 != PyObject_GetBuffer(reader, &result->buffer, PyBUF_CONTIG_RO)) {
+			goto except;
+		}
+	}
+	else {
+		PyErr_SetString(PyExc_ValueError,
+			"must pass an object with a read() method or conforms to buffer protocol");
+		goto except;
+	}
+
+	result->decompressor = self;
+	Py_INCREF(result->decompressor);
+
+	result->inSize = inSize;
+	result->outSize = outSize;
+	result->skipBytes = skipBytes;
+
+	if (ensure_dctx(self, 1)) {
+		goto except;
+	}
+
+	result->input.src = PyMem_Malloc(inSize);
+	if (!result->input.src) {
+		PyErr_NoMemory();
+		goto except;
+	}
+
+	goto finally;
+
+except:
+	Py_CLEAR(result);
+
+finally:
+
+	return result;
+}
+
+PyDoc_STRVAR(Decompressor_stream_reader__doc__,
+"stream_reader(source, [read_size=default, [read_across_frames=False]])\n"
+"\n"
+"Obtain an object that behaves like an I/O stream that can be used for\n"
+"reading decompressed output from an object.\n"
+"\n"
+"The source object can be any object with a ``read(size)`` method or that\n"
+"conforms to the buffer protocol.\n"
+"\n"
+"``read_across_frames`` controls the behavior of ``read()`` when the end\n"
+"of a zstd frame is reached. When ``True``, ``read()`` can potentially\n"
+"return data belonging to multiple zstd frames. When ``False``, ``read()``\n"
+"will return when the end of a frame is reached.\n"
+);
+
+static ZstdDecompressionReader* Decompressor_stream_reader(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"source",
+		"read_size",
+		"read_across_frames",
+		NULL
+	};
+
+	PyObject* source;
+	size_t readSize = ZSTD_DStreamInSize();
+	PyObject* readAcrossFrames = NULL;
+	ZstdDecompressionReader* result;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kO:stream_reader", kwlist,
+		&source, &readSize, &readAcrossFrames)) {
+		return NULL;
+	}
+
+	if (ensure_dctx(self, 1)) {
+		return NULL;
+	}
+
+	result = (ZstdDecompressionReader*)PyObject_CallObject((PyObject*)&ZstdDecompressionReaderType, NULL);
+	if (NULL == result) {
+		return NULL;
+	}
+
+	result->entered = 0;
+	result->closed = 0;
+
+	if (PyObject_HasAttrString(source, "read")) {
+		result->reader = source;
+		Py_INCREF(source);
+		result->readSize = readSize;
+	}
+	else if (1 == PyObject_CheckBuffer(source)) {
+		if (0 != PyObject_GetBuffer(source, &result->buffer, PyBUF_CONTIG_RO)) {
+			Py_CLEAR(result);
+			return NULL;
+		}
+	}
+	else {
+		PyErr_SetString(PyExc_TypeError,
+			"must pass an object with a read() method or that conforms to the buffer protocol");
+		Py_CLEAR(result);
+		return NULL;
+	}
+
+	result->decompressor = self;
+	Py_INCREF(self);
+	result->readAcrossFrames = readAcrossFrames ? PyObject_IsTrue(readAcrossFrames) : 0;
+
+	return result;
+}
+
+PyDoc_STRVAR(Decompressor_stream_writer__doc__,
+"Create a context manager to write decompressed data to an object.\n"
+"\n"
+"The passed object must have a ``write()`` method.\n"
+"\n"
+"The caller feeds intput data to the object by calling ``write(data)``.\n"
+"Decompressed data is written to the argument given as it is decompressed.\n"
+"\n"
+"An optional ``write_size`` argument defines the size of chunks to\n"
+"``write()`` to the writer. It defaults to the default output size for a zstd\n"
+"streaming decompressor.\n"
+);
+
+static ZstdDecompressionWriter* Decompressor_stream_writer(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"writer",
+		"write_size",
+		"write_return_read",
+		NULL
+	};
+
+	PyObject* writer;
+	size_t outSize = ZSTD_DStreamOutSize();
+	PyObject* writeReturnRead = NULL;
+	ZstdDecompressionWriter* result;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kO:stream_writer", kwlist,
+		&writer, &outSize, &writeReturnRead)) {
+		return NULL;
+	}
+
+	if (!PyObject_HasAttrString(writer, "write")) {
+		PyErr_SetString(PyExc_ValueError, "must pass an object with a write() method");
+		return NULL;
+	}
+
+	if (ensure_dctx(self, 1)) {
+		return NULL;
+	}
+
+	result = (ZstdDecompressionWriter*)PyObject_CallObject((PyObject*)&ZstdDecompressionWriterType, NULL);
+	if (!result) {
+		return NULL;
+	}
+
+	result->entered = 0;
+	result->closed = 0;
+
+	result->decompressor = self;
+	Py_INCREF(result->decompressor);
+
+	result->writer = writer;
+	Py_INCREF(result->writer);
+
+	result->outSize = outSize;
+	result->writeReturnRead = writeReturnRead ? PyObject_IsTrue(writeReturnRead) : 0;
+
+	return result;
+}
+
+PyDoc_STRVAR(Decompressor_decompress_content_dict_chain__doc__,
+"Decompress a series of chunks using the content dictionary chaining technique\n"
+);
+
+static PyObject* Decompressor_decompress_content_dict_chain(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"frames",
+		NULL
+	};
+
+	PyObject* chunks;
+	Py_ssize_t chunksLen;
+	Py_ssize_t chunkIndex;
+	char parity = 0;
+	PyObject* chunk;
+	char* chunkData;
+	Py_ssize_t chunkSize;
+	size_t zresult;
+	ZSTD_frameHeader frameHeader;
+	void* buffer1 = NULL;
+	size_t buffer1Size = 0;
+	size_t buffer1ContentSize = 0;
+	void* buffer2 = NULL;
+	size_t buffer2Size = 0;
+	size_t buffer2ContentSize = 0;
+	void* destBuffer = NULL;
+	PyObject* result = NULL;
+	ZSTD_outBuffer outBuffer;
+	ZSTD_inBuffer inBuffer;
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O!:decompress_content_dict_chain",
+		kwlist, &PyList_Type, &chunks)) {
+		return NULL;
+	}
+
+	chunksLen = PyList_Size(chunks);
+	if (!chunksLen) {
+		PyErr_SetString(PyExc_ValueError, "empty input chain");
+		return NULL;
+	}
+
+	/* The first chunk should not be using a dictionary. We handle it specially. */
+	chunk = PyList_GetItem(chunks, 0);
+	if (!PyBytes_Check(chunk)) {
+		PyErr_SetString(PyExc_ValueError, "chunk 0 must be bytes");
+		return NULL;
+	}
+
+	/* We require that all chunks be zstd frames and that they have content size set. */
+	PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
+	zresult = ZSTD_getFrameHeader(&frameHeader, (void*)chunkData, chunkSize);
+	if (ZSTD_isError(zresult)) {
+		PyErr_SetString(PyExc_ValueError, "chunk 0 is not a valid zstd frame");
+		return NULL;
+	}
+	else if (zresult) {
+		PyErr_SetString(PyExc_ValueError, "chunk 0 is too small to contain a zstd frame");
+		return NULL;
+	}
+
+	if (ZSTD_CONTENTSIZE_UNKNOWN == frameHeader.frameContentSize) {
+		PyErr_SetString(PyExc_ValueError, "chunk 0 missing content size in frame");
+		return NULL;
+	}
+
+	assert(ZSTD_CONTENTSIZE_ERROR != frameHeader.frameContentSize);
+
+	/* We check against PY_SSIZE_T_MAX here because we ultimately cast the
+	 * result to a Python object and it's length can be no greater than
+	 * Py_ssize_t. In theory, we could have an intermediate frame that is 
+	 * larger. But a) why would this API be used for frames that large b)
+	 * it isn't worth the complexity to support. */
+	assert(SIZE_MAX >= PY_SSIZE_T_MAX);
+	if (frameHeader.frameContentSize > PY_SSIZE_T_MAX) {
+		PyErr_SetString(PyExc_ValueError,
+			"chunk 0 is too large to decompress on this platform");
+		return NULL;
+	}
+
+	if (ensure_dctx(self, 0)) {
+		goto finally;
+	}
+
+	buffer1Size = (size_t)frameHeader.frameContentSize;
+	buffer1 = PyMem_Malloc(buffer1Size);
+	if (!buffer1) {
+		goto finally;
+	}
+
+	outBuffer.dst = buffer1;
+	outBuffer.size = buffer1Size;
+	outBuffer.pos = 0;
+
+	inBuffer.src = chunkData;
+	inBuffer.size = chunkSize;
+	inBuffer.pos = 0;
+
+	Py_BEGIN_ALLOW_THREADS
+	zresult = ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+	Py_END_ALLOW_THREADS
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "could not decompress chunk 0: %s", ZSTD_getErrorName(zresult));
+		goto finally;
+	}
+	else if (zresult) {
+		PyErr_Format(ZstdError, "chunk 0 did not decompress full frame");
+		goto finally;
+	}
+
+	buffer1ContentSize = outBuffer.pos;
+
+	/* Special case of a simple chain. */
+	if (1 == chunksLen) {
+		result = PyBytes_FromStringAndSize(buffer1, buffer1Size);
+		goto finally;
+	}
+
+	/* This should ideally look at next chunk. But this is slightly simpler. */
+	buffer2Size = (size_t)frameHeader.frameContentSize;
+	buffer2 = PyMem_Malloc(buffer2Size);
+	if (!buffer2) {
+		goto finally;
+	}
+
+	/* For each subsequent chunk, use the previous fulltext as a content dictionary.
+	   Our strategy is to have 2 buffers. One holds the previous fulltext (to be
+	   used as a content dictionary) and the other holds the new fulltext. The
+	   buffers grow when needed but never decrease in size. This limits the
+	   memory allocator overhead.
+	*/
+	for (chunkIndex = 1; chunkIndex < chunksLen; chunkIndex++) {
+		chunk = PyList_GetItem(chunks, chunkIndex);
+		if (!PyBytes_Check(chunk)) {
+			PyErr_Format(PyExc_ValueError, "chunk %zd must be bytes", chunkIndex);
+			goto finally;
+		}
+
+		PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
+		zresult = ZSTD_getFrameHeader(&frameHeader, (void*)chunkData, chunkSize);
+		if (ZSTD_isError(zresult)) {
+			PyErr_Format(PyExc_ValueError, "chunk %zd is not a valid zstd frame", chunkIndex);
+			goto finally;
+		}
+		else if (zresult) {
+			PyErr_Format(PyExc_ValueError, "chunk %zd is too small to contain a zstd frame", chunkIndex);
+			goto finally;
+		}
+
+		if (ZSTD_CONTENTSIZE_UNKNOWN == frameHeader.frameContentSize) {
+			PyErr_Format(PyExc_ValueError, "chunk %zd missing content size in frame", chunkIndex);
+			goto finally;
+		}
+
+		assert(ZSTD_CONTENTSIZE_ERROR != frameHeader.frameContentSize);
+
+		if (frameHeader.frameContentSize > PY_SSIZE_T_MAX) {
+			PyErr_Format(PyExc_ValueError,
+				"chunk %zd is too large to decompress on this platform", chunkIndex);
+			goto finally;
+		}
+
+		inBuffer.src = chunkData;
+		inBuffer.size = chunkSize;
+		inBuffer.pos = 0;
+
+		parity = chunkIndex % 2;
+
+		/* This could definitely be abstracted to reduce code duplication. */
+		if (parity) {
+			/* Resize destination buffer to hold larger content. */
+			if (buffer2Size < frameHeader.frameContentSize) {
+				buffer2Size = (size_t)frameHeader.frameContentSize;
+				destBuffer = PyMem_Realloc(buffer2, buffer2Size);
+				if (!destBuffer) {
+					goto finally;
+				}
+				buffer2 = destBuffer;
+			}
+
+			Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_DCtx_refPrefix_advanced(self->dctx,
+				buffer1, buffer1ContentSize, ZSTD_dct_rawContent);
+			Py_END_ALLOW_THREADS
+			if (ZSTD_isError(zresult)) {
+				PyErr_Format(ZstdError,
+					"failed to load prefix dictionary at chunk %zd", chunkIndex);
+				goto finally;
+			}
+
+			outBuffer.dst = buffer2;
+			outBuffer.size = buffer2Size;
+			outBuffer.pos = 0;
+
+			Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+			Py_END_ALLOW_THREADS
+			if (ZSTD_isError(zresult)) {
+				PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
+					chunkIndex, ZSTD_getErrorName(zresult));
+				goto finally;
+			}
+			else if (zresult) {
+				PyErr_Format(ZstdError, "chunk %zd did not decompress full frame",
+					chunkIndex);
+				goto finally;
+			}
+
+			buffer2ContentSize = outBuffer.pos;
+		}
+		else {
+			if (buffer1Size < frameHeader.frameContentSize) {
+				buffer1Size = (size_t)frameHeader.frameContentSize;
+				destBuffer = PyMem_Realloc(buffer1, buffer1Size);
+				if (!destBuffer) {
+					goto finally;
+				}
+				buffer1 = destBuffer;
+			}
+
+			Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_DCtx_refPrefix_advanced(self->dctx,
+				buffer2, buffer2ContentSize, ZSTD_dct_rawContent);
+			Py_END_ALLOW_THREADS
+			if (ZSTD_isError(zresult)) {
+				PyErr_Format(ZstdError,
+					"failed to load prefix dictionary at chunk %zd", chunkIndex);
+				goto finally;
+			}
+
+			outBuffer.dst = buffer1;
+			outBuffer.size = buffer1Size;
+			outBuffer.pos = 0;
+
+			Py_BEGIN_ALLOW_THREADS
+			zresult = ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+			Py_END_ALLOW_THREADS
+			if (ZSTD_isError(zresult)) {
+				PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
+					chunkIndex, ZSTD_getErrorName(zresult));
+				goto finally;
+			}
+			else if (zresult) {
+				PyErr_Format(ZstdError, "chunk %zd did not decompress full frame",
+					chunkIndex);
+				goto finally;
+			}
+
+			buffer1ContentSize = outBuffer.pos;
+		}
+	}
+
+	result = PyBytes_FromStringAndSize(parity ? buffer2 : buffer1,
+		parity ? buffer2ContentSize : buffer1ContentSize);
+
+finally:
+	if (buffer2) {
+		PyMem_Free(buffer2);
+	}
+	if (buffer1) {
+		PyMem_Free(buffer1);
+	}
+
+	return result;
+}
+
+typedef struct {
+	void* sourceData;
+	size_t sourceSize;
+	size_t destSize;
+} FramePointer;
+
+typedef struct {
+	FramePointer* frames;
+	Py_ssize_t framesSize;
+	unsigned long long compressedSize;
+} FrameSources;
+
+typedef struct {
+	void* dest;
+	Py_ssize_t destSize;
+	BufferSegment* segments;
+	Py_ssize_t segmentsSize;
+} DestBuffer;
+
+typedef enum {
+	WorkerError_none = 0,
+	WorkerError_zstd = 1,
+	WorkerError_memory = 2,
+	WorkerError_sizeMismatch = 3,
+	WorkerError_unknownSize = 4,
+} WorkerError;
+
+typedef struct {
+	/* Source records and length */
+	FramePointer* framePointers;
+	/* Which records to process. */
+	Py_ssize_t startOffset;
+	Py_ssize_t endOffset;
+	unsigned long long totalSourceSize;
+
+	/* Compression state and settings. */
+	ZSTD_DCtx* dctx;
+	int requireOutputSizes;
+
+	/* Output storage. */
+	DestBuffer* destBuffers;
+	Py_ssize_t destCount;
+
+	/* Item that error occurred on. */
+	Py_ssize_t errorOffset;
+	/* If an error occurred. */
+	WorkerError error;
+	/* result from zstd decompression operation */
+	size_t zresult;
+} WorkerState;
+
+static void decompress_worker(WorkerState* state) {
+	size_t allocationSize;
+	DestBuffer* destBuffer;
+	Py_ssize_t frameIndex;
+	Py_ssize_t localOffset = 0;
+	Py_ssize_t currentBufferStartIndex = state->startOffset;
+	Py_ssize_t remainingItems = state->endOffset - state->startOffset + 1;
+	void* tmpBuf;
+	Py_ssize_t destOffset = 0;
+	FramePointer* framePointers = state->framePointers;
+	size_t zresult;
+	unsigned long long totalOutputSize = 0;
+
+	assert(NULL == state->destBuffers);
+	assert(0 == state->destCount);
+	assert(state->endOffset - state->startOffset >= 0);
+
+	/* We could get here due to the way work is allocated. Ideally we wouldn't
+	   get here. But that would require a bit of a refactor in the caller. */
+	if (state->totalSourceSize > SIZE_MAX) {
+		state->error = WorkerError_memory;
+		state->errorOffset = 0;
+		return;
+	}
+
+	/*
+	 * We need to allocate a buffer to hold decompressed data. How we do this
+	 * depends on what we know about the output. The following scenarios are
+	 * possible:
+	 *
+	 * 1. All structs defining frames declare the output size.
+	 * 2. The decompressed size is embedded within the zstd frame.
+	 * 3. The decompressed size is not stored anywhere.
+	 *
+	 * For now, we only support #1 and #2.
+	 */
+
+	/* Resolve ouput segments. */
+	for (frameIndex = state->startOffset; frameIndex <= state->endOffset; frameIndex++) {
+		FramePointer* fp = &framePointers[frameIndex];
+		unsigned long long decompressedSize;
+
+		if (0 == fp->destSize) {
+			decompressedSize = ZSTD_getFrameContentSize(fp->sourceData, fp->sourceSize);
+
+			if (ZSTD_CONTENTSIZE_ERROR == decompressedSize) {
+				state->error = WorkerError_unknownSize;
+				state->errorOffset = frameIndex;
+				return;
+			}
+			else if (ZSTD_CONTENTSIZE_UNKNOWN == decompressedSize) {
+				if (state->requireOutputSizes) {
+					state->error = WorkerError_unknownSize;
+					state->errorOffset = frameIndex;
+					return;
+				}
+
+				/* This will fail the assert for .destSize > 0 below. */
+				decompressedSize = 0;
+			}
+
+			if (decompressedSize > SIZE_MAX) {
+				state->error = WorkerError_memory;
+				state->errorOffset = frameIndex;
+				return;
+			}
+
+			fp->destSize = (size_t)decompressedSize;
+		}
+
+		totalOutputSize += fp->destSize;
+	}
+
+	state->destBuffers = calloc(1, sizeof(DestBuffer));
+	if (NULL == state->destBuffers) {
+		state->error = WorkerError_memory;
+		return;
+	}
+
+	state->destCount = 1;
+
+	destBuffer = &state->destBuffers[state->destCount - 1];
+
+	assert(framePointers[state->startOffset].destSize > 0); /* For now. */
+
+	allocationSize = roundpow2((size_t)state->totalSourceSize);
+
+	if (framePointers[state->startOffset].destSize > allocationSize) {
+		allocationSize = roundpow2(framePointers[state->startOffset].destSize);
+	}
+
+	destBuffer->dest = malloc(allocationSize);
+	if (NULL == destBuffer->dest) {
+		state->error = WorkerError_memory;
+		return;
+	}
+
+	destBuffer->destSize = allocationSize;
+
+	destBuffer->segments = calloc(remainingItems, sizeof(BufferSegment));
+	if (NULL == destBuffer->segments) {
+		/* Caller will free state->dest as part of cleanup. */
+		state->error = WorkerError_memory;
+		return;
+	}
+
+	destBuffer->segmentsSize = remainingItems;
+
+	for (frameIndex = state->startOffset; frameIndex <= state->endOffset; frameIndex++) {
+		ZSTD_outBuffer outBuffer;
+		ZSTD_inBuffer inBuffer;
+		const void* source = framePointers[frameIndex].sourceData;
+		const size_t sourceSize = framePointers[frameIndex].sourceSize;
+		void* dest;
+		const size_t decompressedSize = framePointers[frameIndex].destSize;
+		size_t destAvailable = destBuffer->destSize - destOffset;
+
+		assert(decompressedSize > 0); /* For now. */
+
+		/*
+		 * Not enough space in current buffer. Finish current before and allocate and
+		 * switch to a new one.
+		 */
+		if (decompressedSize > destAvailable) {
+			/*
+			 * Shrinking the destination buffer is optional. But it should be cheap,
+			 * so we just do it.
+			 */
+			if (destAvailable) {
+				tmpBuf = realloc(destBuffer->dest, destOffset);
+				if (NULL == tmpBuf) {
+					state->error = WorkerError_memory;
+					return;
+				}
+
+				destBuffer->dest = tmpBuf;
+				destBuffer->destSize = destOffset;
+			}
+
+			/* Truncate segments buffer. */
+			tmpBuf = realloc(destBuffer->segments,
+				(frameIndex - currentBufferStartIndex) * sizeof(BufferSegment));
+			if (NULL == tmpBuf) {
+				state->error = WorkerError_memory;
+				return;
+			}
+
+			destBuffer->segments = tmpBuf;
+			destBuffer->segmentsSize = frameIndex - currentBufferStartIndex;
+
+			/* Grow space for new DestBuffer. */
+			tmpBuf = realloc(state->destBuffers, (state->destCount + 1) * sizeof(DestBuffer));
+			if (NULL == tmpBuf) {
+				state->error = WorkerError_memory;
+				return;
+			}
+
+			state->destBuffers = tmpBuf;
+			state->destCount++;
+
+			destBuffer = &state->destBuffers[state->destCount - 1];
+
+			/* Don't take any chances will non-NULL pointers. */
+			memset(destBuffer, 0, sizeof(DestBuffer));
+
+			allocationSize = roundpow2((size_t)state->totalSourceSize);
+
+			if (decompressedSize > allocationSize) {
+				allocationSize = roundpow2(decompressedSize);
+			}
+
+			destBuffer->dest = malloc(allocationSize);
+			if (NULL == destBuffer->dest) {
+				state->error = WorkerError_memory;
+				return;
+			}
+
+			destBuffer->destSize = allocationSize;
+			destAvailable = allocationSize;
+			destOffset = 0;
+			localOffset = 0;
+
+			destBuffer->segments = calloc(remainingItems, sizeof(BufferSegment));
+			if (NULL == destBuffer->segments) {
+				state->error = WorkerError_memory;
+				return;
+			}
+
+			destBuffer->segmentsSize = remainingItems;
+			currentBufferStartIndex = frameIndex;
+		}
+
+		dest = (char*)destBuffer->dest + destOffset;
+
+		outBuffer.dst = dest;
+		outBuffer.size = decompressedSize;
+		outBuffer.pos = 0;
+
+		inBuffer.src = source;
+		inBuffer.size = sourceSize;
+		inBuffer.pos = 0;
+
+		zresult = ZSTD_decompressStream(state->dctx, &outBuffer, &inBuffer);
+		if (ZSTD_isError(zresult)) {
+			state->error = WorkerError_zstd;
+			state->zresult = zresult;
+			state->errorOffset = frameIndex;
+			return;
+		}
+		else if (zresult || outBuffer.pos != decompressedSize) {
+			state->error = WorkerError_sizeMismatch;
+			state->zresult = outBuffer.pos;
+			state->errorOffset = frameIndex;
+			return;
+		}
+
+		destBuffer->segments[localOffset].offset = destOffset;
+		destBuffer->segments[localOffset].length = outBuffer.pos;
+		destOffset += outBuffer.pos;
+		localOffset++;
+		remainingItems--;
+	}
+
+	if (destBuffer->destSize > destOffset) {
+		tmpBuf = realloc(destBuffer->dest, destOffset);
+		if (NULL == tmpBuf) {
+			state->error = WorkerError_memory;
+			return;
+		}
+
+		destBuffer->dest = tmpBuf;
+		destBuffer->destSize = destOffset;
+	}
+}
+
+ZstdBufferWithSegmentsCollection* decompress_from_framesources(ZstdDecompressor* decompressor, FrameSources* frames,
+	Py_ssize_t threadCount) {
+	Py_ssize_t i = 0;
+	int errored = 0;
+	Py_ssize_t segmentsCount;
+	ZstdBufferWithSegments* bws = NULL;
+	PyObject* resultArg = NULL;
+	Py_ssize_t resultIndex;
+	ZstdBufferWithSegmentsCollection* result = NULL;
+	FramePointer* framePointers = frames->frames;
+	unsigned long long workerBytes = 0;
+	Py_ssize_t currentThread = 0;
+	Py_ssize_t workerStartOffset = 0;
+	POOL_ctx* pool = NULL;
+	WorkerState* workerStates = NULL;
+	unsigned long long bytesPerWorker;
+
+	/* Caller should normalize 0 and negative values to 1 or larger. */
+	assert(threadCount >= 1);
+
+	/* More threads than inputs makes no sense under any conditions. */
+	threadCount = frames->framesSize < threadCount ? frames->framesSize
+												   : threadCount;
+
+	/* TODO lower thread count if input size is too small and threads would just
+	   add overhead. */
+
+	if (decompressor->dict) {
+		if (ensure_ddict(decompressor->dict)) {
+			return NULL;
+		}
+	}
+
+	/* If threadCount==1, we don't start a thread pool. But we do leverage the
+	   same API for dispatching work. */
+	workerStates = PyMem_Malloc(threadCount * sizeof(WorkerState));
+	if (NULL == workerStates) {
+		PyErr_NoMemory();
+		goto finally;
+	}
+
+	memset(workerStates, 0, threadCount * sizeof(WorkerState));
+
+	if (threadCount > 1) {
+		pool = POOL_create(threadCount, 1);
+		if (NULL == pool) {
+			PyErr_SetString(ZstdError, "could not initialize zstd thread pool");
+			goto finally;
+		}
+	}
+
+	bytesPerWorker = frames->compressedSize / threadCount;
+
+	if (bytesPerWorker > SIZE_MAX) {
+		PyErr_SetString(ZstdError, "too much data per worker for this platform");
+		goto finally;
+	}
+
+	for (i = 0; i < threadCount; i++) {
+		size_t zresult;
+
+		workerStates[i].dctx = ZSTD_createDCtx();
+		if (NULL == workerStates[i].dctx) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		ZSTD_copyDCtx(workerStates[i].dctx, decompressor->dctx);
+
+		if (decompressor->dict) {
+			zresult = ZSTD_DCtx_refDDict(workerStates[i].dctx, decompressor->dict->ddict);
+			if (zresult) {
+				PyErr_Format(ZstdError, "unable to reference prepared dictionary: %s",
+					ZSTD_getErrorName(zresult));
+				goto finally;
+			}
+		}
+
+		workerStates[i].framePointers = framePointers;
+		workerStates[i].requireOutputSizes = 1;
+	}
+
+	Py_BEGIN_ALLOW_THREADS
+	/* There are many ways to split work among workers.
+
+	   For now, we take a simple approach of splitting work so each worker
+	   gets roughly the same number of input bytes. This will result in more
+	   starvation than running N>threadCount jobs. But it avoids complications
+	   around state tracking, which could involve extra locking.
+	*/
+	for (i = 0; i < frames->framesSize; i++) {
+		workerBytes += frames->frames[i].sourceSize;
+
+		/*
+		 * The last worker/thread needs to handle all remaining work. Don't
+		 * trigger it prematurely. Defer to the block outside of the loop.
+		 * (But still process this loop so workerBytes is correct.
+		 */
+		if (currentThread == threadCount - 1) {
+			continue;
+		}
+
+		if (workerBytes >= bytesPerWorker) {
+			workerStates[currentThread].startOffset = workerStartOffset;
+			workerStates[currentThread].endOffset = i;
+			workerStates[currentThread].totalSourceSize = workerBytes;
+
+			if (threadCount > 1) {
+				POOL_add(pool, (POOL_function)decompress_worker, &workerStates[currentThread]);
+			}
+			else {
+				decompress_worker(&workerStates[currentThread]);
+			}
+			currentThread++;
+			workerStartOffset = i + 1;
+			workerBytes = 0;
+		}
+	}
+
+	if (workerBytes) {
+		workerStates[currentThread].startOffset = workerStartOffset;
+		workerStates[currentThread].endOffset = frames->framesSize - 1;
+		workerStates[currentThread].totalSourceSize = workerBytes;
+
+		if (threadCount > 1) {
+			POOL_add(pool, (POOL_function)decompress_worker, &workerStates[currentThread]);
+		}
+		else {
+			decompress_worker(&workerStates[currentThread]);
+		}
+	}
+
+	if (threadCount > 1) {
+		POOL_free(pool);
+		pool = NULL;
+	}
+	Py_END_ALLOW_THREADS
+
+	for (i = 0; i < threadCount; i++) {
+		switch (workerStates[i].error) {
+		case WorkerError_none:
+			break;
+
+		case WorkerError_zstd:
+			PyErr_Format(ZstdError, "error decompressing item %zd: %s",
+				workerStates[i].errorOffset, ZSTD_getErrorName(workerStates[i].zresult));
+			errored = 1;
+			break;
+
+		case WorkerError_memory:
+			PyErr_NoMemory();
+			errored = 1;
+			break;
+
+		case WorkerError_sizeMismatch:
+			PyErr_Format(ZstdError, "error decompressing item %zd: decompressed %zu bytes; expected %zu",
+				workerStates[i].errorOffset, workerStates[i].zresult,
+				framePointers[workerStates[i].errorOffset].destSize);
+			errored = 1;
+			break;
+
+		case WorkerError_unknownSize:
+			PyErr_Format(PyExc_ValueError, "could not determine decompressed size of item %zd",
+				workerStates[i].errorOffset);
+			errored = 1;
+			break;
+
+		default:
+			PyErr_Format(ZstdError, "unhandled error type: %d; this is a bug",
+				workerStates[i].error);
+			errored = 1;
+			break;
+		}
+
+		if (errored) {
+			break;
+		}
+	}
+
+	if (errored) {
+		goto finally;
+	}
+
+	segmentsCount = 0;
+	for (i = 0; i < threadCount; i++) {
+		segmentsCount += workerStates[i].destCount;
+	}
+
+	resultArg = PyTuple_New(segmentsCount);
+	if (NULL == resultArg) {
+		goto finally;
+	}
+
+	resultIndex = 0;
+
+	for (i = 0; i < threadCount; i++) {
+		Py_ssize_t bufferIndex;
+		WorkerState* state = &workerStates[i];
+
+		for (bufferIndex = 0; bufferIndex < state->destCount; bufferIndex++) {
+			DestBuffer* destBuffer = &state->destBuffers[bufferIndex];
+
+			bws = BufferWithSegments_FromMemory(destBuffer->dest, destBuffer->destSize,
+				destBuffer->segments, destBuffer->segmentsSize);
+			if (NULL == bws) {
+				goto finally;
+			}
+
+			/*
+			* Memory for buffer and segments was allocated using malloc() in worker
+			* and the memory is transferred to the BufferWithSegments instance. So
+			* tell instance to use free() and NULL the reference in the state struct
+			* so it isn't freed below.
+			*/
+			bws->useFree = 1;
+			destBuffer->dest = NULL;
+			destBuffer->segments = NULL;
+
+			PyTuple_SET_ITEM(resultArg, resultIndex++, (PyObject*)bws);
+		}
+	}
+
+	result = (ZstdBufferWithSegmentsCollection*)PyObject_CallObject(
+		(PyObject*)&ZstdBufferWithSegmentsCollectionType, resultArg);
+
+finally:
+	Py_CLEAR(resultArg);
+
+	if (workerStates) {
+		for (i = 0; i < threadCount; i++) {
+			Py_ssize_t bufferIndex;
+			WorkerState* state = &workerStates[i];
+
+			if (state->dctx) {
+				ZSTD_freeDCtx(state->dctx);
+			}
+
+			for (bufferIndex = 0; bufferIndex < state->destCount; bufferIndex++) {
+				if (state->destBuffers) {
+					/*
+					* Will be NULL if memory transfered to a BufferWithSegments.
+					* Otherwise it is left over after an error occurred.
+					*/
+					free(state->destBuffers[bufferIndex].dest);
+					free(state->destBuffers[bufferIndex].segments);
+				}
+			}
+
+			free(state->destBuffers);
+		}
+
+		PyMem_Free(workerStates);
+	}
+
+	POOL_free(pool);
+
+	return result;
+}
+
+PyDoc_STRVAR(Decompressor_multi_decompress_to_buffer__doc__,
+"Decompress multiple frames to output buffers\n"
+"\n"
+"Receives a ``BufferWithSegments``, a ``BufferWithSegmentsCollection`` or a\n"
+"list of bytes-like objects. Each item in the passed collection should be a\n"
+"compressed zstd frame.\n"
+"\n"
+"Unless ``decompressed_sizes`` is specified, the content size *must* be\n"
+"written into the zstd frame header. If ``decompressed_sizes`` is specified,\n"
+"it is an object conforming to the buffer protocol that represents an array\n"
+"of 64-bit unsigned integers in the machine's native format. Specifying\n"
+"``decompressed_sizes`` avoids a pre-scan of each frame to determine its\n"
+"output size.\n"
+"\n"
+"Returns a ``BufferWithSegmentsCollection`` containing the decompressed\n"
+"data. All decompressed data is allocated in a single memory buffer. The\n"
+"``BufferWithSegments`` instance tracks which objects are at which offsets\n"
+"and their respective lengths.\n"
+"\n"
+"The ``threads`` argument controls how many threads to use for operations.\n"
+"Negative values will use the same number of threads as logical CPUs on the\n"
+"machine.\n"
+);
+
+static ZstdBufferWithSegmentsCollection* Decompressor_multi_decompress_to_buffer(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"frames",
+		"decompressed_sizes",
+		"threads",
+		NULL
+	};
+
+	PyObject* frames;
+	Py_buffer frameSizes;
+	int threads = 0;
+	Py_ssize_t frameCount;
+	Py_buffer* frameBuffers = NULL;
+	FramePointer* framePointers = NULL;
+	unsigned long long* frameSizesP = NULL;
+	unsigned long long totalInputSize = 0;
+	FrameSources frameSources;
+	ZstdBufferWithSegmentsCollection* result = NULL;
+	Py_ssize_t i;
+
+	memset(&frameSizes, 0, sizeof(frameSizes));
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|y*i:multi_decompress_to_buffer",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|s*i:multi_decompress_to_buffer",
+#endif
+		kwlist, &frames, &frameSizes, &threads)) {
+		return NULL;
+	}
+
+	if (frameSizes.buf) {
+		if (!PyBuffer_IsContiguous(&frameSizes, 'C') || frameSizes.ndim > 1) {
+			PyErr_SetString(PyExc_ValueError, "decompressed_sizes buffer should be contiguous and have a single dimension");
+			goto finally;
+		}
+
+		frameSizesP = (unsigned long long*)frameSizes.buf;
+	}
+
+	if (threads < 0) {
+		threads = cpu_count();
+	}
+
+	if (threads < 2) {
+		threads = 1;
+	}
+
+	if (PyObject_TypeCheck(frames, &ZstdBufferWithSegmentsType)) {
+		ZstdBufferWithSegments* buffer = (ZstdBufferWithSegments*)frames;
+		frameCount = buffer->segmentCount;
+
+		if (frameSizes.buf && frameSizes.len != frameCount * (Py_ssize_t)sizeof(unsigned long long)) {
+			PyErr_Format(PyExc_ValueError, "decompressed_sizes size mismatch; expected %zd, got %zd",
+				frameCount * sizeof(unsigned long long), frameSizes.len);
+			goto finally;
+		}
+
+		framePointers = PyMem_Malloc(frameCount * sizeof(FramePointer));
+		if (!framePointers) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		for (i = 0; i < frameCount; i++) {
+			void* sourceData;
+			unsigned long long sourceSize;
+			unsigned long long decompressedSize = 0;
+
+			if (buffer->segments[i].offset + buffer->segments[i].length > buffer->dataSize) {
+				PyErr_Format(PyExc_ValueError, "item %zd has offset outside memory area", i);
+				goto finally;
+			}
+
+			sourceData = (char*)buffer->data + buffer->segments[i].offset;
+			sourceSize = buffer->segments[i].length;
+			totalInputSize += sourceSize;
+
+			if (frameSizesP) {
+				decompressedSize = frameSizesP[i];
+			}
+
+			if (sourceSize > SIZE_MAX) {
+				PyErr_Format(PyExc_ValueError,
+					"item %zd is too large for this platform", i);
+				goto finally;
+			}
+
+			if (decompressedSize > SIZE_MAX) {
+				PyErr_Format(PyExc_ValueError,
+					"decompressed size of item %zd is too large for this platform", i);
+				goto finally;
+			}
+
+			framePointers[i].sourceData = sourceData;
+			framePointers[i].sourceSize = (size_t)sourceSize;
+			framePointers[i].destSize = (size_t)decompressedSize;
+		}
+	}
+	else if (PyObject_TypeCheck(frames, &ZstdBufferWithSegmentsCollectionType)) {
+		Py_ssize_t offset = 0;
+		ZstdBufferWithSegments* buffer;
+		ZstdBufferWithSegmentsCollection* collection = (ZstdBufferWithSegmentsCollection*)frames;
+
+		frameCount = BufferWithSegmentsCollection_length(collection);
+
+		if (frameSizes.buf && frameSizes.len != frameCount) {
+			PyErr_Format(PyExc_ValueError,
+				"decompressed_sizes size mismatch; expected %zd; got %zd",
+				frameCount * sizeof(unsigned long long), frameSizes.len);
+			goto finally;
+		}
+
+		framePointers = PyMem_Malloc(frameCount * sizeof(FramePointer));
+		if (NULL == framePointers) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		/* Iterate the data structure directly because it is faster. */
+		for (i = 0; i < collection->bufferCount; i++) {
+			Py_ssize_t segmentIndex;
+			buffer = collection->buffers[i];
+
+			for (segmentIndex = 0; segmentIndex < buffer->segmentCount; segmentIndex++) {
+				unsigned long long decompressedSize = frameSizesP ? frameSizesP[offset] : 0;
+
+				if (buffer->segments[segmentIndex].offset + buffer->segments[segmentIndex].length > buffer->dataSize) {
+					PyErr_Format(PyExc_ValueError, "item %zd has offset outside memory area",
+						offset);
+					goto finally;
+				}
+
+				if (buffer->segments[segmentIndex].length > SIZE_MAX) {
+					PyErr_Format(PyExc_ValueError,
+						"item %zd in buffer %zd is too large for this platform",
+						segmentIndex, i);
+					goto finally;
+				}
+
+				if (decompressedSize > SIZE_MAX) {
+					PyErr_Format(PyExc_ValueError,
+						"decompressed size of item %zd in buffer %zd is too large for this platform",
+						segmentIndex, i);
+					goto finally;
+				}
+
+				totalInputSize += buffer->segments[segmentIndex].length;
+
+				framePointers[offset].sourceData = (char*)buffer->data + buffer->segments[segmentIndex].offset;
+				framePointers[offset].sourceSize = (size_t)buffer->segments[segmentIndex].length;
+				framePointers[offset].destSize = (size_t)decompressedSize;
+
+				offset++;
+			}
+		}
+	}
+	else if (PyList_Check(frames)) {
+		frameCount = PyList_GET_SIZE(frames);
+
+		if (frameSizes.buf && frameSizes.len != frameCount * (Py_ssize_t)sizeof(unsigned long long)) {
+			PyErr_Format(PyExc_ValueError, "decompressed_sizes size mismatch; expected %zd, got %zd",
+				frameCount * sizeof(unsigned long long), frameSizes.len);
+			goto finally;
+		}
+
+		framePointers = PyMem_Malloc(frameCount * sizeof(FramePointer));
+		if (!framePointers) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		frameBuffers = PyMem_Malloc(frameCount * sizeof(Py_buffer));
+		if (NULL == frameBuffers) {
+			PyErr_NoMemory();
+			goto finally;
+		}
+
+		memset(frameBuffers, 0, frameCount * sizeof(Py_buffer));
+
+		/* Do a pass to assemble info about our input buffers and output sizes. */
+		for (i = 0; i < frameCount; i++) {
+			unsigned long long decompressedSize = frameSizesP ? frameSizesP[i] : 0;
+
+			if (0 != PyObject_GetBuffer(PyList_GET_ITEM(frames, i),
+				&frameBuffers[i], PyBUF_CONTIG_RO)) {
+				PyErr_Clear();
+				PyErr_Format(PyExc_TypeError, "item %zd not a bytes like object", i);
+				goto finally;
+			}
+
+			if (decompressedSize > SIZE_MAX) {
+				PyErr_Format(PyExc_ValueError,
+					"decompressed size of item %zd is too large for this platform", i);
+				goto finally;
+			}
+
+			totalInputSize += frameBuffers[i].len;
+
+			framePointers[i].sourceData = frameBuffers[i].buf;
+			framePointers[i].sourceSize = frameBuffers[i].len;
+			framePointers[i].destSize = (size_t)decompressedSize;
+		}
+	}
+	else {
+		PyErr_SetString(PyExc_TypeError, "argument must be list or BufferWithSegments");
+		goto finally;
+	}
+
+	/* We now have an array with info about our inputs and outputs. Feed it into
+	   our generic decompression function. */
+	frameSources.frames = framePointers;
+	frameSources.framesSize = frameCount;
+	frameSources.compressedSize = totalInputSize;
+
+	result = decompress_from_framesources(self, &frameSources, threads);
+
+finally:
+	if (frameSizes.buf) {
+		PyBuffer_Release(&frameSizes);
+	}
+	PyMem_Free(framePointers);
+
+	if (frameBuffers) {
+		for (i = 0; i < frameCount; i++) {
+			PyBuffer_Release(&frameBuffers[i]);
+		}
+
+		PyMem_Free(frameBuffers);
+	}
+
+	return result;
+}
+
+static PyMethodDef Decompressor_methods[] = {
+	{ "copy_stream", (PyCFunction)Decompressor_copy_stream, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_copy_stream__doc__ },
+	{ "decompress", (PyCFunction)Decompressor_decompress, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_decompress__doc__ },
+	{ "decompressobj", (PyCFunction)Decompressor_decompressobj, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_decompressobj__doc__ },
+	{ "read_to_iter", (PyCFunction)Decompressor_read_to_iter, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_read_to_iter__doc__ },
+	/* TODO Remove deprecated API */
+	{ "read_from", (PyCFunction)Decompressor_read_to_iter, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_read_to_iter__doc__ },
+	{ "stream_reader", (PyCFunction)Decompressor_stream_reader,
+	METH_VARARGS | METH_KEYWORDS, Decompressor_stream_reader__doc__ },
+	{ "stream_writer", (PyCFunction)Decompressor_stream_writer, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_stream_writer__doc__ },
+	/* TODO remove deprecated API */
+	{ "write_to", (PyCFunction)Decompressor_stream_writer, METH_VARARGS | METH_KEYWORDS,
+	Decompressor_stream_writer__doc__ },
+	{ "decompress_content_dict_chain", (PyCFunction)Decompressor_decompress_content_dict_chain,
+	  METH_VARARGS | METH_KEYWORDS, Decompressor_decompress_content_dict_chain__doc__ },
+	{ "multi_decompress_to_buffer", (PyCFunction)Decompressor_multi_decompress_to_buffer,
+	  METH_VARARGS | METH_KEYWORDS, Decompressor_multi_decompress_to_buffer__doc__ },
+	{ "memory_size", (PyCFunction)Decompressor_memory_size, METH_NOARGS,
+	Decompressor_memory_size__doc__ },
+	{ NULL, NULL }
+};
+
+PyTypeObject ZstdDecompressorType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdDecompressor",        /* tp_name */
+	sizeof(ZstdDecompressor),       /* tp_basicsize */
+	0,                              /* tp_itemsize */
+	(destructor)Decompressor_dealloc, /* tp_dealloc */
+	0,                              /* tp_print */
+	0,                              /* tp_getattr */
+	0,                              /* tp_setattr */
+	0,                              /* tp_compare */
+	0,                              /* tp_repr */
+	0,                              /* tp_as_number */
+	0,                              /* tp_as_sequence */
+	0,                              /* tp_as_mapping */
+	0,                              /* tp_hash */
+	0,                              /* tp_call */
+	0,                              /* tp_str */
+	0,                              /* tp_getattro */
+	0,                              /* tp_setattro */
+	0,                              /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	Decompressor__doc__,            /* tp_doc */
+	0,                              /* tp_traverse */
+	0,                              /* tp_clear */
+	0,                              /* tp_richcompare */
+	0,                              /* tp_weaklistoffset */
+	0,                              /* tp_iter */
+	0,                              /* tp_iternext */
+	Decompressor_methods,           /* tp_methods */
+	0,                              /* tp_members */
+	0,                              /* tp_getset */
+	0,                              /* tp_base */
+	0,                              /* tp_dict */
+	0,                              /* tp_descr_get */
+	0,                              /* tp_descr_set */
+	0,                              /* tp_dictoffset */
+	(initproc)Decompressor_init,    /* tp_init */
+	0,                              /* tp_alloc */
+	PyType_GenericNew,              /* tp_new */
+};
+
+void decompressor_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdDecompressorType) = &PyType_Type;
+	if (PyType_Ready(&ZstdDecompressorType) < 0) {
+		return;
+	}
+
+	Py_INCREF((PyObject*)&ZstdDecompressorType);
+	PyModule_AddObject(mod, "ZstdDecompressor",
+		(PyObject*)&ZstdDecompressorType);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/decompressoriterator.c b/contrib/python/zstandard/py2/c-ext/decompressoriterator.c
new file mode 100644
index 00000000000..54b56581de1
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/decompressoriterator.c
@@ -0,0 +1,249 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+#define min(a, b) (((a) < (b)) ? (a) : (b))
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(ZstdDecompressorIterator__doc__,
+"Represents an iterator of decompressed data.\n"
+);
+
+static void ZstdDecompressorIterator_dealloc(ZstdDecompressorIterator* self) {
+	Py_XDECREF(self->decompressor);
+	Py_XDECREF(self->reader);
+
+	if (self->buffer.buf) {
+		PyBuffer_Release(&self->buffer);
+		memset(&self->buffer, 0, sizeof(self->buffer));
+	}
+
+	if (self->input.src) {
+		PyMem_Free((void*)self->input.src);
+		self->input.src = NULL;
+	}
+
+	PyObject_Del(self);
+}
+
+static PyObject* ZstdDecompressorIterator_iter(PyObject* self) {
+	Py_INCREF(self);
+	return self;
+}
+
+static DecompressorIteratorResult read_decompressor_iterator(ZstdDecompressorIterator* self) {
+	size_t zresult;
+	PyObject* chunk;
+	DecompressorIteratorResult result;
+	size_t oldInputPos = self->input.pos;
+
+	result.chunk = NULL;
+
+	chunk = PyBytes_FromStringAndSize(NULL, self->outSize);
+	if (!chunk) {
+		result.errored = 1;
+		return result;
+	}
+
+	self->output.dst = PyBytes_AsString(chunk);
+	self->output.size = self->outSize;
+	self->output.pos = 0;
+
+	Py_BEGIN_ALLOW_THREADS
+	zresult = ZSTD_decompressStream(self->decompressor->dctx, &self->output, &self->input);
+	Py_END_ALLOW_THREADS
+
+	/* We're done with the pointer. Nullify to prevent anyone from getting a
+	handle on a Python object. */
+	self->output.dst = NULL;
+
+	if (ZSTD_isError(zresult)) {
+		Py_DECREF(chunk);
+		PyErr_Format(ZstdError, "zstd decompress error: %s",
+			ZSTD_getErrorName(zresult));
+		result.errored = 1;
+		return result;
+	}
+
+	self->readCount += self->input.pos - oldInputPos;
+
+	/* Frame is fully decoded. Input exhausted and output sitting in buffer. */
+	if (0 == zresult) {
+		self->finishedInput = 1;
+		self->finishedOutput = 1;
+	}
+
+	/* If it produced output data, return it. */
+	if (self->output.pos) {
+		if (self->output.pos < self->outSize) {
+			if (safe_pybytes_resize(&chunk, self->output.pos)) {
+				Py_XDECREF(chunk);
+				result.errored = 1;
+				return result;
+			}
+		}
+	}
+	else {
+		Py_DECREF(chunk);
+		chunk = NULL;
+	}
+
+	result.errored = 0;
+	result.chunk = chunk;
+
+	return result;
+}
+
+static PyObject* ZstdDecompressorIterator_iternext(ZstdDecompressorIterator* self) {
+	PyObject* readResult = NULL;
+	char* readBuffer;
+	Py_ssize_t readSize;
+	Py_ssize_t bufferRemaining;
+	DecompressorIteratorResult result;
+
+	if (self->finishedOutput) {
+		PyErr_SetString(PyExc_StopIteration, "output flushed");
+		return NULL;
+	}
+
+	/* If we have data left in the input, consume it. */
+	if (self->input.pos < self->input.size) {
+		result = read_decompressor_iterator(self);
+		if (result.chunk || result.errored) {
+			return result.chunk;
+		}
+
+		/* Else fall through to get more data from input. */
+	}
+
+read_from_source:
+
+	if (!self->finishedInput) {
+		if (self->reader) {
+			readResult = PyObject_CallMethod(self->reader, "read", "I", self->inSize);
+			if (!readResult) {
+				return NULL;
+			}
+
+			PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+		}
+		else {
+			assert(self->buffer.buf);
+
+			/* Only support contiguous C arrays for now */
+			assert(self->buffer.strides == NULL && self->buffer.suboffsets == NULL);
+			assert(self->buffer.itemsize == 1);
+
+			/* TODO avoid memcpy() below */
+			readBuffer = (char *)self->buffer.buf + self->bufferOffset;
+			bufferRemaining = self->buffer.len - self->bufferOffset;
+			readSize = min(bufferRemaining, (Py_ssize_t)self->inSize);
+			self->bufferOffset += readSize;
+		}
+
+		if (readSize) {
+			if (!self->readCount && self->skipBytes) {
+				assert(self->skipBytes < self->inSize);
+				if ((Py_ssize_t)self->skipBytes >= readSize) {
+					PyErr_SetString(PyExc_ValueError,
+						"skip_bytes larger than first input chunk; "
+						"this scenario is currently unsupported");
+					Py_XDECREF(readResult);
+					return NULL;
+				}
+
+				readBuffer = readBuffer + self->skipBytes;
+				readSize -= self->skipBytes;
+			}
+
+			/* Copy input into previously allocated buffer because it can live longer
+			than a single function call and we don't want to keep a ref to a Python
+			object around. This could be changed... */
+			memcpy((void*)self->input.src, readBuffer, readSize);
+			self->input.size = readSize;
+			self->input.pos = 0;
+		}
+		/* No bytes on first read must mean an empty input stream. */
+		else if (!self->readCount) {
+			self->finishedInput = 1;
+			self->finishedOutput = 1;
+			Py_XDECREF(readResult);
+			PyErr_SetString(PyExc_StopIteration, "empty input");
+			return NULL;
+		}
+		else {
+			self->finishedInput = 1;
+		}
+
+		/* We've copied the data managed by memory. Discard the Python object. */
+		Py_XDECREF(readResult);
+	}
+
+	result = read_decompressor_iterator(self);
+	if (result.errored || result.chunk) {
+		return result.chunk;
+	}
+
+	/* No new output data. Try again unless we know there is no more data. */
+	if (!self->finishedInput) {
+		goto read_from_source;
+	}
+
+	PyErr_SetString(PyExc_StopIteration, "input exhausted");
+	return NULL;
+}
+
+PyTypeObject ZstdDecompressorIteratorType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"zstd.ZstdDecompressorIterator",   /* tp_name */
+	sizeof(ZstdDecompressorIterator),  /* tp_basicsize */
+	0,                                 /* tp_itemsize */
+	(destructor)ZstdDecompressorIterator_dealloc, /* tp_dealloc */
+	0,                                 /* tp_print */
+	0,                                 /* tp_getattr */
+	0,                                 /* tp_setattr */
+	0,                                 /* tp_compare */
+	0,                                 /* tp_repr */
+	0,                                 /* tp_as_number */
+	0,                                 /* tp_as_sequence */
+	0,                                 /* tp_as_mapping */
+	0,                                 /* tp_hash */
+	0,                                 /* tp_call */
+	0,                                 /* tp_str */
+	0,                                 /* tp_getattro */
+	0,                                 /* tp_setattro */
+	0,                                 /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE, /* tp_flags */
+	ZstdDecompressorIterator__doc__,   /* tp_doc */
+	0,                                 /* tp_traverse */
+	0,                                 /* tp_clear */
+	0,                                 /* tp_richcompare */
+	0,                                 /* tp_weaklistoffset */
+	ZstdDecompressorIterator_iter,     /* tp_iter */
+	(iternextfunc)ZstdDecompressorIterator_iternext, /* tp_iternext */
+	0,                                 /* tp_methods */
+	0,                                 /* tp_members */
+	0,                                 /* tp_getset */
+	0,                                 /* tp_base */
+	0,                                 /* tp_dict */
+	0,                                 /* tp_descr_get */
+	0,                                 /* tp_descr_set */
+	0,                                 /* tp_dictoffset */
+	0,                                 /* tp_init */
+	0,                                 /* tp_alloc */
+	PyType_GenericNew,                 /* tp_new */
+};
+
+void decompressoriterator_module_init(PyObject* mod) {
+	Py_TYPE(&ZstdDecompressorIteratorType) = &PyType_Type;
+	if (PyType_Ready(&ZstdDecompressorIteratorType) < 0) {
+		return;
+	}
+}
diff --git a/contrib/python/zstandard/py2/c-ext/frameparams.c b/contrib/python/zstandard/py2/c-ext/frameparams.c
new file mode 100644
index 00000000000..35ca3ca5900
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/frameparams.c
@@ -0,0 +1,138 @@
+/**
+* Copyright (c) 2017-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#include "python-zstandard.h"
+
+extern PyObject* ZstdError;
+
+PyDoc_STRVAR(FrameParameters__doc__,
+	"FrameParameters: information about a zstd frame");
+
+FrameParametersObject* get_frame_parameters(PyObject* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"data",
+		NULL
+	};
+
+	Py_buffer source;
+	ZSTD_frameHeader header;
+	FrameParametersObject* result = NULL;
+	size_t zresult;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:get_frame_parameters",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:get_frame_parameters",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	zresult = ZSTD_getFrameHeader(&header, source.buf, source.len);
+
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "cannot get frame parameters: %s", ZSTD_getErrorName(zresult));
+		goto finally;
+	}
+
+	if (zresult) {
+		PyErr_Format(ZstdError, "not enough data for frame parameters; need %zu bytes", zresult);
+		goto finally;
+	}
+
+	result = PyObject_New(FrameParametersObject, &FrameParametersType);
+	if (!result) {
+		goto finally;
+	}
+
+	result->frameContentSize = header.frameContentSize;
+	result->windowSize = header.windowSize;
+	result->dictID = header.dictID;
+	result->checksumFlag = header.checksumFlag ? 1 : 0;
+
+finally:
+	PyBuffer_Release(&source);
+	return result;
+}
+
+static void FrameParameters_dealloc(PyObject* self) {
+	PyObject_Del(self);
+}
+
+static PyMemberDef FrameParameters_members[] = {
+	{ "content_size", T_ULONGLONG,
+	  offsetof(FrameParametersObject, frameContentSize), READONLY,
+	  "frame content size" },
+	{ "window_size", T_ULONGLONG,
+	  offsetof(FrameParametersObject, windowSize), READONLY,
+	  "window size" },
+	{ "dict_id", T_UINT,
+	  offsetof(FrameParametersObject, dictID), READONLY,
+	  "dictionary ID" },
+	{ "has_checksum", T_BOOL,
+	  offsetof(FrameParametersObject, checksumFlag), READONLY,
+	  "checksum flag" },
+	{ NULL }
+};
+
+PyTypeObject FrameParametersType = {
+	PyVarObject_HEAD_INIT(NULL, 0)
+	"FrameParameters",          /* tp_name */
+	sizeof(FrameParametersObject), /* tp_basicsize */
+	0,                         /* tp_itemsize */
+	(destructor)FrameParameters_dealloc, /* tp_dealloc */
+	0,                         /* tp_print */
+	0,                         /* tp_getattr */
+	0,                         /* tp_setattr */
+	0,                         /* tp_compare */
+	0,                         /* tp_repr */
+	0,                         /* tp_as_number */
+	0,                         /* tp_as_sequence */
+	0,                         /* tp_as_mapping */
+	0,                         /* tp_hash  */
+	0,                         /* tp_call */
+	0,                         /* tp_str */
+	0,                         /* tp_getattro */
+	0,                         /* tp_setattro */
+	0,                         /* tp_as_buffer */
+	Py_TPFLAGS_DEFAULT,        /* tp_flags */
+	FrameParameters__doc__,    /* tp_doc */
+	0,                         /* tp_traverse */
+	0,                         /* tp_clear */
+	0,                         /* tp_richcompare */
+	0,                         /* tp_weaklistoffset */
+	0,                         /* tp_iter */
+	0,                         /* tp_iternext */
+	0,                         /* tp_methods */
+	FrameParameters_members,   /* tp_members */
+	0,                         /* tp_getset */
+	0,                         /* tp_base */
+	0,                         /* tp_dict */
+	0,                         /* tp_descr_get */
+	0,                         /* tp_descr_set */
+	0,                         /* tp_dictoffset */
+	0,                         /* tp_init */
+	0,                         /* tp_alloc */
+	0,                         /* tp_new */
+};
+
+void frameparams_module_init(PyObject* mod) {
+	Py_TYPE(&FrameParametersType) = &PyType_Type;
+	if (PyType_Ready(&FrameParametersType) < 0) {
+		return;
+	}
+
+	Py_INCREF(&FrameParametersType);
+	PyModule_AddObject(mod, "FrameParameters", (PyObject*)&FrameParametersType);
+}
diff --git a/contrib/python/zstandard/py2/c-ext/python-zstandard.h b/contrib/python/zstandard/py2/c-ext/python-zstandard.h
new file mode 100644
index 00000000000..bd1cb4dcad8
--- /dev/null
+++ b/contrib/python/zstandard/py2/c-ext/python-zstandard.h
@@ -0,0 +1,359 @@
+/**
+* Copyright (c) 2016-present, Gregory Szorc
+* All rights reserved.
+*
+* This software may be modified and distributed under the terms
+* of the BSD license. See the LICENSE file for details.
+*/
+
+#define PY_SSIZE_T_CLEAN
+#include <Python.h>
+#include "structmember.h"
+
+#define ZSTD_STATIC_LINKING_ONLY
+#define ZDICT_STATIC_LINKING_ONLY
+#include <zstd.h>
+#include <zdict.h>
+
+/* Remember to change the string in zstandard/__init__ as well */
+#define PYTHON_ZSTANDARD_VERSION "0.14.1"
+
+typedef enum {
+	compressorobj_flush_finish,
+	compressorobj_flush_block,
+} CompressorObj_Flush;
+
+/*
+   Represents a ZstdCompressionParameters type.
+
+   This type holds all the low-level compression parameters that can be set.
+*/
+typedef struct {
+	PyObject_HEAD
+	ZSTD_CCtx_params* params;
+} ZstdCompressionParametersObject;
+
+extern PyTypeObject ZstdCompressionParametersType;
+
+/*
+   Represents a FrameParameters type.
+
+   This type is basically a wrapper around ZSTD_frameParams.
+*/
+typedef struct {
+	PyObject_HEAD
+	unsigned long long frameContentSize;
+	unsigned long long windowSize;
+	unsigned dictID;
+	char checksumFlag;
+} FrameParametersObject;
+
+extern PyTypeObject FrameParametersType;
+
+/*
+   Represents a ZstdCompressionDict type.
+
+   Instances hold data used for a zstd compression dictionary.
+*/
+typedef struct {
+	PyObject_HEAD
+
+	/* Pointer to dictionary data. Owned by self. */
+	void* dictData;
+	/* Size of dictionary data. */
+	size_t dictSize;
+	ZSTD_dictContentType_e dictType;
+	/* k parameter for cover dictionaries. Only populated by train_cover_dict(). */
+	unsigned k;
+	/* d parameter for cover dictionaries. Only populated by train_cover_dict(). */
+	unsigned d;
+	/* Digested dictionary, suitable for reuse. */
+	ZSTD_CDict* cdict;
+	ZSTD_DDict* ddict;
+} ZstdCompressionDict;
+
+extern PyTypeObject ZstdCompressionDictType;
+
+/*
+   Represents a ZstdCompressor type.
+*/
+typedef struct {
+	PyObject_HEAD
+
+	/* Number of threads to use for operations. */
+	unsigned int threads;
+	/* Pointer to compression dictionary to use. NULL if not using dictionary
+	   compression. */
+	ZstdCompressionDict* dict;
+	/* Compression context to use. Populated during object construction. */
+	ZSTD_CCtx* cctx;
+	/* Compression parameters in use. */
+	ZSTD_CCtx_params* params;
+} ZstdCompressor;
+
+extern PyTypeObject ZstdCompressorType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdCompressor* compressor;
+	ZSTD_outBuffer output;
+	int finished;
+} ZstdCompressionObj;
+
+extern PyTypeObject ZstdCompressionObjType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdCompressor* compressor;
+	PyObject* writer;
+	ZSTD_outBuffer output;
+	size_t outSize;
+	int entered;
+	int closed;
+	int writeReturnRead;
+	unsigned long long bytesCompressed;
+} ZstdCompressionWriter;
+
+extern PyTypeObject ZstdCompressionWriterType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdCompressor* compressor;
+	PyObject* reader;
+	Py_buffer buffer;
+	Py_ssize_t bufferOffset;
+	size_t inSize;
+	size_t outSize;
+
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	int finishedOutput;
+	int finishedInput;
+	PyObject* readResult;
+} ZstdCompressorIterator;
+
+extern PyTypeObject ZstdCompressorIteratorType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdCompressor* compressor;
+	PyObject* reader;
+	Py_buffer buffer;
+	size_t readSize;
+
+	int entered;
+	int closed;
+	unsigned long long bytesCompressed;
+
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	int finishedInput;
+	int finishedOutput;
+	PyObject* readResult;
+} ZstdCompressionReader;
+
+extern PyTypeObject ZstdCompressionReaderType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdCompressor* compressor;
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	Py_buffer inBuffer;
+	int finished;
+	size_t chunkSize;
+} ZstdCompressionChunker;
+
+extern PyTypeObject ZstdCompressionChunkerType;
+
+typedef enum {
+	compressionchunker_mode_normal,
+	compressionchunker_mode_flush,
+	compressionchunker_mode_finish,
+} CompressionChunkerMode;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdCompressionChunker* chunker;
+	CompressionChunkerMode mode;
+} ZstdCompressionChunkerIterator;
+
+extern PyTypeObject ZstdCompressionChunkerIteratorType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZSTD_DCtx* dctx;
+	ZstdCompressionDict* dict;
+	size_t maxWindowSize;
+	ZSTD_format_e format;
+} ZstdDecompressor;
+
+extern PyTypeObject ZstdDecompressorType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdDecompressor* decompressor;
+	size_t outSize;
+	int finished;
+} ZstdDecompressionObj;
+
+extern PyTypeObject ZstdDecompressionObjType;
+
+typedef struct {
+	PyObject_HEAD
+
+	/* Parent decompressor to which this object is associated. */
+	ZstdDecompressor* decompressor;
+	/* Object to read() from (if reading from a stream). */
+	PyObject* reader;
+	/* Size for read() operations on reader. */
+	size_t readSize;
+	/* Whether a read() can return data spanning multiple zstd frames. */
+	int readAcrossFrames;
+	/* Buffer to read from (if reading from a buffer). */
+	Py_buffer buffer;
+
+	/* Whether the context manager is active. */
+	int entered;
+	/* Whether we've closed the stream. */
+	int closed;
+
+	/* Number of bytes decompressed and returned to user. */
+	unsigned long long bytesDecompressed;
+
+	/* Tracks data going into decompressor. */
+	ZSTD_inBuffer input;
+
+	/* Holds output from read() operation on reader. */
+	PyObject* readResult;
+
+	/* Whether all input has been sent to the decompressor. */
+	int finishedInput;
+	/* Whether all output has been flushed from the decompressor. */
+	int finishedOutput;
+} ZstdDecompressionReader;
+
+extern PyTypeObject ZstdDecompressionReaderType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdDecompressor* decompressor;
+	PyObject* writer;
+	size_t outSize;
+	int entered;
+	int closed;
+	int writeReturnRead;
+} ZstdDecompressionWriter;
+
+extern PyTypeObject ZstdDecompressionWriterType;
+
+typedef struct {
+	PyObject_HEAD
+
+	ZstdDecompressor* decompressor;
+	PyObject* reader;
+	Py_buffer buffer;
+	Py_ssize_t bufferOffset;
+	size_t inSize;
+	size_t outSize;
+	size_t skipBytes;
+	ZSTD_inBuffer input;
+	ZSTD_outBuffer output;
+	Py_ssize_t readCount;
+	int finishedInput;
+	int finishedOutput;
+} ZstdDecompressorIterator;
+
+extern PyTypeObject ZstdDecompressorIteratorType;
+
+typedef struct {
+	int errored;
+	PyObject* chunk;
+} DecompressorIteratorResult;
+
+typedef struct {
+	/* The public API is that these are 64-bit unsigned integers. So these can't
+	 * be size_t, even though values larger than SIZE_MAX or PY_SSIZE_T_MAX may 
+	 * be nonsensical for this platform. */
+	unsigned long long offset;
+	unsigned long long length;
+} BufferSegment;
+
+typedef struct {
+	PyObject_HEAD
+
+	PyObject* parent;
+	BufferSegment* segments;
+	Py_ssize_t segmentCount;
+} ZstdBufferSegments;
+
+extern PyTypeObject ZstdBufferSegmentsType;
+
+typedef struct {
+	PyObject_HEAD
+
+	PyObject* parent;
+	void* data;
+	Py_ssize_t dataSize;
+	unsigned long long offset;
+} ZstdBufferSegment;
+
+extern PyTypeObject ZstdBufferSegmentType;
+
+typedef struct {
+	PyObject_HEAD
+
+	Py_buffer parent;
+	void* data;
+	unsigned long long dataSize;
+	BufferSegment* segments;
+	Py_ssize_t segmentCount;
+	int useFree;
+} ZstdBufferWithSegments;
+
+extern PyTypeObject ZstdBufferWithSegmentsType;
+
+/**
+ * An ordered collection of BufferWithSegments exposed as a squashed collection.
+ *
+ * This type provides a virtual view spanning multiple BufferWithSegments
+ * instances. It allows multiple instances to be "chained" together and
+ * exposed as a single collection. e.g. if there are 2 buffers holding
+ * 10 segments each, then o[14] will access the 5th segment in the 2nd buffer.
+ */
+typedef struct {
+	PyObject_HEAD
+
+	/* An array of buffers that should be exposed through this instance. */
+	ZstdBufferWithSegments** buffers;
+	/* Number of elements in buffers array. */
+	Py_ssize_t bufferCount;
+	/* Array of first offset in each buffer instance. 0th entry corresponds
+	   to number of elements in the 0th buffer. 1st entry corresponds to the
+	   sum of elements in 0th and 1st buffers. */
+	Py_ssize_t* firstElements;
+} ZstdBufferWithSegmentsCollection;
+
+extern PyTypeObject ZstdBufferWithSegmentsCollectionType;
+
+int set_parameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int value);
+int set_parameters(ZSTD_CCtx_params* params, ZstdCompressionParametersObject* obj);
+int to_cparams(ZstdCompressionParametersObject* params, ZSTD_compressionParameters* cparams);
+FrameParametersObject* get_frame_parameters(PyObject* self, PyObject* args, PyObject* kwargs);
+int ensure_ddict(ZstdCompressionDict* dict);
+int ensure_dctx(ZstdDecompressor* decompressor, int loadDict);
+ZstdCompressionDict* train_dictionary(PyObject* self, PyObject* args, PyObject* kwargs);
+ZstdBufferWithSegments* BufferWithSegments_FromMemory(void* data, unsigned long long dataSize, BufferSegment* segments, Py_ssize_t segmentsSize);
+Py_ssize_t BufferWithSegmentsCollection_length(ZstdBufferWithSegmentsCollection*);
+int cpu_count(void);
+size_t roundpow2(size_t);
+int safe_pybytes_resize(PyObject** obj, Py_ssize_t size);
diff --git a/contrib/python/zstandard/py2/ya.make b/contrib/python/zstandard/py2/ya.make
new file mode 100644
index 00000000000..2efc543fe85
--- /dev/null
+++ b/contrib/python/zstandard/py2/ya.make
@@ -0,0 +1,58 @@
+# Generated by devtools/yamaker (pypi).
+
+PY2_LIBRARY()
+
+VERSION(0.14.1)
+
+LICENSE(BSD-3-Clause)
+
+PEERDIR(
+    contrib/libs/zstd
+)
+
+ADDINCL(
+    contrib/libs/zstd/include
+    contrib/libs/zstd/lib/common
+    contrib/python/zstandard/py2/c-ext
+)
+
+NO_COMPILER_WARNINGS()
+
+NO_LINT()
+
+SRCS(
+    c-ext/bufferutil.c
+    c-ext/compressionchunker.c
+    c-ext/compressiondict.c
+    c-ext/compressionparams.c
+    c-ext/compressionreader.c
+    c-ext/compressionwriter.c
+    c-ext/compressobj.c
+    c-ext/compressor.c
+    c-ext/compressoriterator.c
+    c-ext/constants.c
+    c-ext/decompressionreader.c
+    c-ext/decompressionwriter.c
+    c-ext/decompressobj.c
+    c-ext/decompressor.c
+    c-ext/decompressoriterator.c
+    c-ext/frameparams.c
+    zstd.c
+)
+
+PY_REGISTER(
+    zstd
+)
+
+PY_SRCS(
+    TOP_LEVEL
+    zstandard/__init__.py
+)
+
+RESOURCE_FILES(
+    PREFIX contrib/python/zstandard/py2/
+    .dist-info/METADATA
+    .dist-info/top_level.txt
+)
+
+END()
diff --git a/contrib/python/zstandard/py2/zstandard/__init__.py b/contrib/python/zstandard/py2/zstandard/__init__.py
new file mode 100644
index 00000000000..5b0a9318761
--- /dev/null
+++ b/contrib/python/zstandard/py2/zstandard/__init__.py
@@ -0,0 +1,75 @@
+# Copyright (c) 2017-present, Gregory Szorc
+# All rights reserved.
+#
+# This software may be modified and distributed under the terms
+# of the BSD license. See the LICENSE file for details.
+
+"""Python interface to the Zstandard (zstd) compression library."""
+
+from __future__ import absolute_import, unicode_literals
+
+# This module serves 2 roles:
+#
+# 1) Export the C or CFFI "backend" through a central module.
+# 2) Implement additional functionality built on top of C or CFFI backend.
+
+import os
+import platform
+
+# Some Python implementations don't support C extensions. That's why we have
+# a CFFI implementation in the first place. The code here import one of our
+# "backends" then re-exports the symbols from this module. For convenience,
+# we support falling back to the CFFI backend if the C extension can't be
+# imported. But for performance reasons, we only do this on unknown Python
+# implementation. Notably, for CPython we require the C extension by default.
+# Because someone will inevitably want special behavior, the behavior is
+# configurable via an environment variable. A potentially better way to handle
+# this is to import a special ``__importpolicy__`` module or something
+# defining a variable and `setup.py` could write the file with whatever
+# policy was specified at build time. Until someone needs it, we go with
+# the hacky but simple environment variable approach.
+_module_policy = os.environ.get("PYTHON_ZSTANDARD_IMPORT_POLICY", "default")
+
+if _module_policy == "default":
+    if platform.python_implementation() in ("CPython",):
+        from zstd import *
+
+        backend = "cext"
+    elif platform.python_implementation() in ("PyPy",):
+        from .cffi import *
+
+        backend = "cffi"
+    else:
+        try:
+            from zstd import *
+
+            backend = "cext"
+        except ImportError:
+            from .cffi import *
+
+            backend = "cffi"
+elif _module_policy == "cffi_fallback":
+    try:
+        from zstd import *
+
+        backend = "cext"
+    except ImportError:
+        from .cffi import *
+
+        backend = "cffi"
+elif _module_policy == "cext":
+    from zstd import *
+
+    backend = "cext"
+elif _module_policy == "cffi":
+    from .cffi import *
+
+    backend = "cffi"
+else:
+    raise ImportError(
+        "unknown module import policy: %s; use default, cffi_fallback, "
+        "cext, or cffi" % _module_policy
+    )
+
+# Keep this in sync with python-zstandard.h.
+__version__ = "0.14.1"
diff --git a/contrib/python/zstandard/py2/zstd.c b/contrib/python/zstandard/py2/zstd.c
new file mode 100644
index 00000000000..3ab69a31139
--- /dev/null
+++ b/contrib/python/zstandard/py2/zstd.c
@@ -0,0 +1,344 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+/* A Python C extension for Zstandard. */
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#elif defined(__APPLE__) || defined(__OpenBSD__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__DragonFly__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#include "python-zstandard.h"
+
+PyObject *ZstdError;
+
+PyDoc_STRVAR(estimate_decompression_context_size__doc__,
+"estimate_decompression_context_size()\n"
+"\n"
+"Estimate the amount of memory allocated to a decompression context.\n"
+);
+
+static PyObject* estimate_decompression_context_size(PyObject* self) {
+	return PyLong_FromSize_t(ZSTD_estimateDCtxSize());
+}
+
+PyDoc_STRVAR(frame_content_size__doc__,
+"frame_content_size(data)\n"
+"\n"
+"Obtain the decompressed size of a frame."
+);
+
+static PyObject* frame_content_size(PyObject* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"source",
+		NULL
+	};
+
+	Py_buffer source;
+	PyObject* result = NULL;
+	unsigned long long size;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:frame_content_size",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:frame_content_size",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	size = ZSTD_getFrameContentSize(source.buf, source.len);
+
+	if (size == ZSTD_CONTENTSIZE_ERROR) {
+		PyErr_SetString(ZstdError, "error when determining content size");
+	}
+	else if (size == ZSTD_CONTENTSIZE_UNKNOWN) {
+		result = PyLong_FromLong(-1);
+	}
+	else {
+		result = PyLong_FromUnsignedLongLong(size);
+	}
+
+finally:
+	PyBuffer_Release(&source);
+
+	return result;
+}
+
+PyDoc_STRVAR(frame_header_size__doc__,
+"frame_header_size(data)\n"
+"\n"
+"Obtain the size of a frame header.\n"
+);
+
+static PyObject* frame_header_size(PyObject* self, PyObject* args, PyObject* kwargs) {
+	static char* kwlist[] = {
+		"source",
+		NULL
+	};
+
+	Py_buffer source;
+	PyObject* result = NULL;
+	size_t zresult;
+
+#if PY_MAJOR_VERSION >= 3
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:frame_header_size",
+#else
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*:frame_header_size",
+#endif
+		kwlist, &source)) {
+		return NULL;
+	}
+
+	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
+		PyErr_SetString(PyExc_ValueError,
+			"data buffer should be contiguous and have at most one dimension");
+		goto finally;
+	}
+
+	zresult = ZSTD_frameHeaderSize(source.buf, source.len);
+	if (ZSTD_isError(zresult)) {
+		PyErr_Format(ZstdError, "could not determine frame header size: %s",
+			ZSTD_getErrorName(zresult));
+	}
+	else {
+		result = PyLong_FromSize_t(zresult);
+	}
+
+finally:
+
+	PyBuffer_Release(&source);
+
+	return result;
+}
+
+PyDoc_STRVAR(get_frame_parameters__doc__,
+"get_frame_parameters(data)\n"
+"\n"
+"Obtains a ``FrameParameters`` instance by parsing data.\n");
+
+PyDoc_STRVAR(train_dictionary__doc__,
+"train_dictionary(dict_size, samples, k=None, d=None, steps=None,\n"
+"                 threads=None,notifications=0, dict_id=0, level=0)\n"
+"\n"
+"Train a dictionary from sample data using the COVER algorithm.\n"
+"\n"
+"A compression dictionary of size ``dict_size`` will be created from the\n"
+"iterable of ``samples``. The raw dictionary bytes will be returned.\n"
+"\n"
+"The COVER algorithm has 2 parameters: ``k`` and ``d``. These control the\n"
+"*segment size* and *dmer size*. A reasonable range for ``k`` is\n"
+"``[16, 2048+]``. A reasonable range for ``d`` is ``[6, 16]``.\n"
+"``d`` must be less than or equal to ``k``.\n"
+"\n"
+"``steps`` can be specified to control the number of steps through potential\n"
+"values of ``k`` and ``d`` to try. ``k`` and ``d`` will only be varied if\n"
+"those arguments are not defined. i.e. if ``d`` is ``8``, then only ``k``\n"
+"will be varied in this mode.\n"
+"\n"
+"``threads`` can specify how many threads to use to test various ``k`` and\n"
+"``d`` values. ``-1`` will use as many threads as available CPUs. By default,\n"
+"a single thread is used.\n"
+"\n"
+"When ``k`` and ``d`` are not defined, default values are used and the\n"
+"algorithm will perform multiple iterations - or steps - to try to find\n"
+"ideal parameters. If both ``k`` and ``d`` are specified, then those values\n"
+"will be used. ``steps`` or ``threads`` triggers optimization mode to test\n"
+"multiple ``k`` and ``d`` variations.\n"
+);
+
+static char zstd_doc[] = "Interface to zstandard";
+
+static PyMethodDef zstd_methods[] = {
+	{ "estimate_decompression_context_size", (PyCFunction)estimate_decompression_context_size,
+	METH_NOARGS, estimate_decompression_context_size__doc__ },
+	{ "frame_content_size", (PyCFunction)frame_content_size,
+	METH_VARARGS | METH_KEYWORDS, frame_content_size__doc__ },
+	{ "frame_header_size", (PyCFunction)frame_header_size,
+	METH_VARARGS | METH_KEYWORDS, frame_header_size__doc__ },
+	{ "get_frame_parameters", (PyCFunction)get_frame_parameters,
+	METH_VARARGS | METH_KEYWORDS, get_frame_parameters__doc__ },
+	{ "train_dictionary", (PyCFunction)train_dictionary,
+	METH_VARARGS | METH_KEYWORDS, train_dictionary__doc__ },
+	{ NULL, NULL }
+};
+
+void bufferutil_module_init(PyObject* mod);
+void compressobj_module_init(PyObject* mod);
+void compressor_module_init(PyObject* mod);
+void compressionparams_module_init(PyObject* mod);
+void constants_module_init(PyObject* mod);
+void compressionchunker_module_init(PyObject* mod);
+void compressiondict_module_init(PyObject* mod);
+void compressionreader_module_init(PyObject* mod);
+void compressionwriter_module_init(PyObject* mod);
+void compressoriterator_module_init(PyObject* mod);
+void decompressor_module_init(PyObject* mod);
+void decompressobj_module_init(PyObject* mod);
+void decompressionreader_module_init(PyObject *mod);
+void decompressionwriter_module_init(PyObject* mod);
+void decompressoriterator_module_init(PyObject* mod);
+void frameparams_module_init(PyObject* mod);
+
+void zstd_module_init(PyObject* m) {
+	/* python-zstandard relies on unstable zstd C API features. This means
+	   that changes in zstd may break expectations in python-zstandard.
+
+	   python-zstandard is distributed with a copy of the zstd sources.
+	   python-zstandard is only guaranteed to work with the bundled version
+	   of zstd.
+
+	   However, downstream redistributors or packagers may unbundle zstd
+	   from python-zstandard. This can result in a mismatch between zstd
+	   versions and API semantics. This essentially "voids the warranty"
+	   of python-zstandard and may cause undefined behavior.
+
+	   We detect this mismatch here and refuse to load the module if this
+	   scenario is detected.
+	*/
+	if (ZSTD_VERSION_NUMBER != 10505 || ZSTD_versionNumber() != 10505) {
+		PyErr_SetString(PyExc_ImportError, "zstd C API mismatch; Python bindings not compiled against expected zstd version");
+		return;
+	}
+
+	bufferutil_module_init(m);
+	compressionparams_module_init(m);
+	compressiondict_module_init(m);
+	compressobj_module_init(m);
+	compressor_module_init(m);
+	compressionchunker_module_init(m);
+	compressionreader_module_init(m);
+	compressionwriter_module_init(m);
+	compressoriterator_module_init(m);
+	constants_module_init(m);
+	decompressor_module_init(m);
+	decompressobj_module_init(m);
+	decompressionreader_module_init(m);
+	decompressionwriter_module_init(m);
+	decompressoriterator_module_init(m);
+	frameparams_module_init(m);
+}
+
+#if defined(__GNUC__) && (__GNUC__ >= 4)
+#  define PYTHON_ZSTD_VISIBILITY __attribute__ ((visibility ("default")))
+#else
+#  define PYTHON_ZSTD_VISIBILITY
+#endif
+
+#if PY_MAJOR_VERSION >= 3
+static struct PyModuleDef zstd_module = {
+	PyModuleDef_HEAD_INIT,
+	"zstd",
+	zstd_doc,
+	-1,
+	zstd_methods
+};
+
+PYTHON_ZSTD_VISIBILITY PyMODINIT_FUNC PyInit_zstd(void) {
+	PyObject *m = PyModule_Create(&zstd_module);
+	if (m) {
+		zstd_module_init(m);
+		if (PyErr_Occurred()) {
+			Py_DECREF(m);
+			m = NULL;
+		}
+	}
+	return m;
+}
+#else
+PYTHON_ZSTD_VISIBILITY PyMODINIT_FUNC initzstd(void) {
+	PyObject *m = Py_InitModule3("zstd", zstd_methods, zstd_doc);
+	if (m) {
+		zstd_module_init(m);
+	}
+}
+#endif
+
+/* Attempt to resolve the number of CPUs in the system. */
+int cpu_count() {
+	int count = 0;
+
+#if defined(_WIN32)
+	SYSTEM_INFO si;
+	si.dwNumberOfProcessors = 0;
+	GetSystemInfo(&si);
+	count = si.dwNumberOfProcessors;
+#elif defined(__APPLE__)
+	int num;
+	size_t size = sizeof(int);
+
+	if (0 == sysctlbyname("hw.logicalcpu", &num, &size, NULL, 0)) {
+		count = num;
+	}
+#elif defined(__linux__)
+	count = sysconf(_SC_NPROCESSORS_ONLN);
+#elif defined(__OpenBSD__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__DragonFly__)
+	int mib[2];
+	size_t len = sizeof(count);
+	mib[0] = CTL_HW;
+	mib[1] = HW_NCPU;
+	if (0 != sysctl(mib, 2, &count, &len, NULL, 0)) {
+		count = 0;
+	}
+#elif defined(__hpux)
+	count = mpctl(MPC_GETNUMSPUS, NULL, NULL);
+#endif
+
+	return count;
+}
+
+size_t roundpow2(size_t i) {
+	i--;
+	i |= i >> 1;
+	i |= i >> 2;
+	i |= i >> 4;
+	i |= i >> 8;
+	i |= i >> 16;
+	i++;
+
+	return i;
+}
+
+/* Safer version of _PyBytes_Resize().
+ *
+ * _PyBytes_Resize() only works if the refcount is 1. In some scenarios,
+ * we can get an object with a refcount > 1, even if it was just created
+ * with PyBytes_FromStringAndSize()! That's because (at least) CPython
+ * pre-allocates PyBytes instances of size 1 for every possible byte value.
+ *
+ * If non-0 is returned, obj may or may not be NULL.
+ */
+int safe_pybytes_resize(PyObject** obj, Py_ssize_t size) {
+	PyObject* tmp;
+
+	if ((*obj)->ob_refcnt == 1) {
+		return _PyBytes_Resize(obj, size);
+	}
+
+	tmp = PyBytes_FromStringAndSize(NULL, size);
+	if (!tmp) {
+		return -1;
+	}
+
+	memcpy(PyBytes_AS_STRING(tmp), PyBytes_AS_STRING(*obj),
+		PyBytes_GET_SIZE(*obj));
+
+	Py_DECREF(*obj);
+	*obj = tmp;
+
+	return 0;
+}
\ No newline at end of file
diff --git a/contrib/python/zstandard/py3/.dist-info/METADATA b/contrib/python/zstandard/py3/.dist-info/METADATA
new file mode 100644
index 00000000000..6eaf2530266
--- /dev/null
+++ b/contrib/python/zstandard/py3/.dist-info/METADATA
@@ -0,0 +1,61 @@
+Metadata-Version: 2.1
+Name: zstandard
+Version: 0.21.0
+Summary: Zstandard bindings for Python
+Home-page: https://github.com/indygreg/python-zstandard
+Author: Gregory Szorc
+Author-email: [email protected]
+License: BSD
+Keywords: zstandard,zstd,compression
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: BSD License
+Classifier: Programming Language :: C
+Classifier: Programming Language :: Python :: 3.7
+Classifier: Programming Language :: Python :: 3.8
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Requires-Python: >=3.7
+License-File: LICENSE
+Requires-Dist: cffi (>=1.11) ; platform_python_implementation == "PyPy"
+Provides-Extra: cffi
+Requires-Dist: cffi (>=1.11) ; extra == 'cffi'
+
+================
+python-zstandard
+================
+
+| |ci-test| |ci-wheel| |ci-typing| |ci-sdist| |ci-anaconda| |ci-sphinx|
+
+This project provides Python bindings for interfacing with the
+`Zstandard <http://www.zstd.net>`_ compression library. A C extension
+and CFFI interface are provided.
+
+The primary goal of the project is to provide a rich interface to the
+underlying C API through a Pythonic interface while not sacrificing
+performance. This means exposing most of the features and flexibility
+of the C API while not sacrificing usability or safety that Python provides.
+
+The canonical home for this project is
+https://github.com/indygreg/python-zstandard.
+
+For usage documentation, see https://python-zstandard.readthedocs.org/.
+
+.. |ci-test| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/test.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/test.yml
+
+.. |ci-wheel| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/wheel.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/wheel.yml
+
+.. |ci-typing| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/typing.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/typing.yml
+
+.. |ci-sdist| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/sdist.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/sdist.yml
+
+.. |ci-anaconda| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/anaconda.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/anaconda.yml
+
+.. |ci-sphinx| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/sphinx.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/sphinx.yml
diff --git a/contrib/python/zstandard/py3/.dist-info/top_level.txt b/contrib/python/zstandard/py3/.dist-info/top_level.txt
new file mode 100644
index 00000000000..864700d2b3e
--- /dev/null
+++ b/contrib/python/zstandard/py3/.dist-info/top_level.txt
@@ -0,0 +1 @@
+zstandard
diff --git a/contrib/python/zstandard/py3/LICENSE b/contrib/python/zstandard/py3/LICENSE
new file mode 100644
index 00000000000..dcec4760996
--- /dev/null
+++ b/contrib/python/zstandard/py3/LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2016, Gregory Szorc
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+this list of conditions and the following disclaimer in the documentation
+and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its contributors
+may be used to endorse or promote products derived from this software without
+specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/contrib/python/zstandard/py3/README.rst b/contrib/python/zstandard/py3/README.rst
new file mode 100644
index 00000000000..bec82d2c3eb
--- /dev/null
+++ b/contrib/python/zstandard/py3/README.rst
@@ -0,0 +1,37 @@
+================
+python-zstandard
+================
+
+| |ci-test| |ci-wheel| |ci-typing| |ci-sdist| |ci-anaconda| |ci-sphinx|
+
+This project provides Python bindings for interfacing with the
+`Zstandard <http://www.zstd.net>`_ compression library. A C extension
+and CFFI interface are provided.
+
+The primary goal of the project is to provide a rich interface to the
+underlying C API through a Pythonic interface while not sacrificing
+performance. This means exposing most of the features and flexibility
+of the C API while not sacrificing usability or safety that Python provides.
+
+The canonical home for this project is
+https://github.com/indygreg/python-zstandard.
+
+For usage documentation, see https://python-zstandard.readthedocs.org/.
+
+.. |ci-test| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/test.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/test.yml
+
+.. |ci-wheel| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/wheel.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/wheel.yml
+
+.. |ci-typing| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/typing.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/typing.yml
+
+.. |ci-sdist| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/sdist.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/sdist.yml
+
+.. |ci-anaconda| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/anaconda.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/anaconda.yml
+
+.. |ci-sphinx| image:: https://github.com/indygreg/python-zstandard/workflows/.github/workflows/sphinx.yml/badge.svg
+     :target: https://github.com/indygreg/python-zstandard/blob/main/.github/workflows/sphinx.yml
diff --git a/contrib/python/zstandard/py3/c-ext/backend_c.c b/contrib/python/zstandard/py3/c-ext/backend_c.c
new file mode 100644
index 00000000000..bf61f9c8488
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/backend_c.c
@@ -0,0 +1,353 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+/* A Python C extension for Zstandard. */
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#elif defined(__APPLE__) || defined(__OpenBSD__) || defined(__FreeBSD__) ||    \
+    defined(__NetBSD__) || defined(__DragonFly__)
+#include <sys/types.h>
+
+#include <sys/sysctl.h>
+
+#endif
+
+#include "python-zstandard.h"
+
+#include "bufferutil.c"
+#include "compressionchunker.c"
+#include "compressiondict.c"
+#include "compressionparams.c"
+#include "compressionreader.c"
+#include "compressionwriter.c"
+#include "compressobj.c"
+#include "compressor.c"
+#include "compressoriterator.c"
+#include "constants.c"
+#include "decompressionreader.c"
+#include "decompressionwriter.c"
+#include "decompressobj.c"
+#include "decompressor.c"
+#include "decompressoriterator.c"
+#include "frameparams.c"
+
+PyObject *ZstdError;
+
+static PyObject *estimate_decompression_context_size(PyObject *self) {
+    return PyLong_FromSize_t(ZSTD_estimateDCtxSize());
+}
+
+static PyObject *frame_content_size(PyObject *self, PyObject *args,
+                                    PyObject *kwargs) {
+    static char *kwlist[] = {"source", NULL};
+
+    Py_buffer source;
+    PyObject *result = NULL;
+    unsigned long long size;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:frame_content_size",
+                                     kwlist, &source)) {
+        return NULL;
+    }
+
+    size = ZSTD_getFrameContentSize(source.buf, source.len);
+
+    if (size == ZSTD_CONTENTSIZE_ERROR) {
+        PyErr_SetString(ZstdError, "error when determining content size");
+    }
+    else if (size == ZSTD_CONTENTSIZE_UNKNOWN) {
+        result = PyLong_FromLong(-1);
+    }
+    else {
+        result = PyLong_FromUnsignedLongLong(size);
+    }
+
+    PyBuffer_Release(&source);
+
+    return result;
+}
+
+static PyObject *frame_header_size(PyObject *self, PyObject *args,
+                                   PyObject *kwargs) {
+    static char *kwlist[] = {"source", NULL};
+
+    Py_buffer source;
+    PyObject *result = NULL;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:frame_header_size",
+                                     kwlist, &source)) {
+        return NULL;
+    }
+
+    zresult = ZSTD_frameHeaderSize(source.buf, source.len);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "could not determine frame header size: %s",
+                     ZSTD_getErrorName(zresult));
+    }
+    else {
+        result = PyLong_FromSize_t(zresult);
+    }
+
+    PyBuffer_Release(&source);
+
+    return result;
+}
+
+static char zstd_doc[] = "Interface to zstandard";
+
+static PyMethodDef zstd_methods[] = {
+    {"estimate_decompression_context_size",
+     (PyCFunction)estimate_decompression_context_size, METH_NOARGS, NULL},
+    {"frame_content_size", (PyCFunction)frame_content_size,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"frame_header_size", (PyCFunction)frame_header_size,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"get_frame_parameters", (PyCFunction)get_frame_parameters,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"train_dictionary", (PyCFunction)train_dictionary,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {NULL, NULL}};
+
+void bufferutil_module_init(PyObject *mod);
+void compressobj_module_init(PyObject *mod);
+void compressor_module_init(PyObject *mod);
+void compressionparams_module_init(PyObject *mod);
+void constants_module_init(PyObject *mod);
+void compressionchunker_module_init(PyObject *mod);
+void compressiondict_module_init(PyObject *mod);
+void compressionreader_module_init(PyObject *mod);
+void compressionwriter_module_init(PyObject *mod);
+void compressoriterator_module_init(PyObject *mod);
+void decompressor_module_init(PyObject *mod);
+void decompressobj_module_init(PyObject *mod);
+void decompressionreader_module_init(PyObject *mod);
+void decompressionwriter_module_init(PyObject *mod);
+void decompressoriterator_module_init(PyObject *mod);
+void frameparams_module_init(PyObject *mod);
+
+void zstd_module_init(PyObject *m) {
+    /* python-zstandard relies on unstable zstd C API features. This means
+       that changes in zstd may break expectations in python-zstandard.
+
+       python-zstandard is distributed with a copy of the zstd sources.
+       python-zstandard is only guaranteed to work with the bundled version
+       of zstd.
+
+       However, downstream redistributors or packagers may unbundle zstd
+       from python-zstandard. This can result in a mismatch between zstd
+       versions and API semantics. This essentially "voids the warranty"
+       of python-zstandard and may cause undefined behavior.
+
+       We detect this mismatch here and refuse to load the module if this
+       scenario is detected.
+    */
+    PyObject *features = NULL;
+    PyObject *feature = NULL;
+    unsigned zstd_ver_no = ZSTD_versionNumber();
+    unsigned our_hardcoded_version = 10505;
+    if (ZSTD_VERSION_NUMBER != our_hardcoded_version ||
+        zstd_ver_no != our_hardcoded_version) {
+        PyErr_Format(
+            PyExc_ImportError,
+            "zstd C API versions mismatch; Python bindings were not "
+            "compiled/linked against expected zstd version (%u returned by the "
+            "lib, %u hardcoded in zstd headers, %u hardcoded in the cext)",
+            zstd_ver_no, ZSTD_VERSION_NUMBER, our_hardcoded_version);
+        return;
+    }
+
+    features = PySet_New(NULL);
+    if (NULL == features) {
+        PyErr_SetString(PyExc_ImportError, "could not create empty set");
+        return;
+    }
+
+    feature = PyUnicode_FromString("buffer_types");
+    if (NULL == feature) {
+        PyErr_SetString(PyExc_ImportError, "could not create feature string");
+        return;
+    }
+
+    if (PySet_Add(features, feature) == -1) {
+        return;
+    }
+
+    Py_DECREF(feature);
+
+#ifdef HAVE_ZSTD_POOL_APIS
+    feature = PyUnicode_FromString("multi_compress_to_buffer");
+    if (NULL == feature) {
+        PyErr_SetString(PyExc_ImportError, "could not create feature string");
+        return;
+    }
+
+    if (PySet_Add(features, feature) == -1) {
+        return;
+    }
+
+    Py_DECREF(feature);
+#endif
+
+#ifdef HAVE_ZSTD_POOL_APIS
+    feature = PyUnicode_FromString("multi_decompress_to_buffer");
+    if (NULL == feature) {
+        PyErr_SetString(PyExc_ImportError, "could not create feature string");
+        return;
+    }
+
+    if (PySet_Add(features, feature) == -1) {
+        return;
+    }
+
+    Py_DECREF(feature);
+#endif
+
+    if (PyObject_SetAttrString(m, "backend_features", features) == -1) {
+        return;
+    }
+
+    Py_DECREF(features);
+
+    bufferutil_module_init(m);
+    compressionparams_module_init(m);
+    compressiondict_module_init(m);
+    compressobj_module_init(m);
+    compressor_module_init(m);
+    compressionchunker_module_init(m);
+    compressionreader_module_init(m);
+    compressionwriter_module_init(m);
+    compressoriterator_module_init(m);
+    constants_module_init(m);
+    decompressor_module_init(m);
+    decompressobj_module_init(m);
+    decompressionreader_module_init(m);
+    decompressionwriter_module_init(m);
+    decompressoriterator_module_init(m);
+    frameparams_module_init(m);
+}
+
+#if defined(__GNUC__) && (__GNUC__ >= 4)
+#define PYTHON_ZSTD_VISIBILITY __attribute__((visibility("default")))
+#else
+#define PYTHON_ZSTD_VISIBILITY
+#endif
+
+static struct PyModuleDef zstd_module = {PyModuleDef_HEAD_INIT, "backend_c",
+                                         zstd_doc, -1, zstd_methods};
+
+PYTHON_ZSTD_VISIBILITY PyMODINIT_FUNC PyInit_backend_c(void) {
+    PyObject *m = PyModule_Create(&zstd_module);
+    if (m) {
+        zstd_module_init(m);
+        if (PyErr_Occurred()) {
+            Py_DECREF(m);
+            m = NULL;
+        }
+    }
+    return m;
+}
+
+/* Attempt to resolve the number of CPUs in the system. */
+int cpu_count() {
+    int count = 0;
+
+#if defined(_WIN32)
+    SYSTEM_INFO si;
+    si.dwNumberOfProcessors = 0;
+    GetSystemInfo(&si);
+    count = si.dwNumberOfProcessors;
+#elif defined(__APPLE__)
+    int num;
+    size_t size = sizeof(int);
+
+    if (0 == sysctlbyname("hw.logicalcpu", &num, &size, NULL, 0)) {
+        count = num;
+    }
+#elif defined(__linux__)
+    count = sysconf(_SC_NPROCESSORS_ONLN);
+#elif defined(__OpenBSD__) || defined(__FreeBSD__) || defined(__NetBSD__) ||   \
+    defined(__DragonFly__)
+    int mib[2];
+    size_t len = sizeof(count);
+    mib[0] = CTL_HW;
+    mib[1] = HW_NCPU;
+    if (0 != sysctl(mib, 2, &count, &len, NULL, 0)) {
+        count = 0;
+    }
+#elif defined(__hpux)
+    count = mpctl(MPC_GETNUMSPUS, NULL, NULL);
+#endif
+
+    return count;
+}
+
+size_t roundpow2(size_t i) {
+    i--;
+    i |= i >> 1;
+    i |= i >> 2;
+    i |= i >> 4;
+    i |= i >> 8;
+    i |= i >> 16;
+    i++;
+
+    return i;
+}
+
+/* Safer version of _PyBytes_Resize().
+ *
+ * _PyBytes_Resize() only works if the refcount is 1. In some scenarios,
+ * we can get an object with a refcount > 1, even if it was just created
+ * with PyBytes_FromStringAndSize()! That's because (at least) CPython
+ * pre-allocates PyBytes instances of size 1 for every possible byte value.
+ *
+ * If non-0 is returned, obj may or may not be NULL.
+ */
+int safe_pybytes_resize(PyObject **obj, Py_ssize_t size) {
+    PyObject *tmp;
+
+    if ((*obj)->ob_refcnt == 1) {
+        return _PyBytes_Resize(obj, size);
+    }
+
+    tmp = PyBytes_FromStringAndSize(NULL, size);
+    if (!tmp) {
+        return -1;
+    }
+
+    memcpy(PyBytes_AS_STRING(tmp), PyBytes_AS_STRING(*obj),
+           PyBytes_GET_SIZE(*obj));
+
+    Py_DECREF(*obj);
+    *obj = tmp;
+
+    return 0;
+}
+
+// Set/raise an `io.UnsupportedOperation` exception.
+void set_io_unsupported_operation(void) {
+    PyObject *iomod;
+    PyObject *exc;
+
+    iomod = PyImport_ImportModule("io");
+    if (NULL == iomod) {
+        return;
+    }
+
+    exc = PyObject_GetAttrString(iomod, "UnsupportedOperation");
+    if (NULL == exc) {
+        Py_DECREF(iomod);
+        return;
+    }
+
+    PyErr_SetNone(exc);
+    Py_DECREF(exc);
+    Py_DECREF(iomod);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/bufferutil.c b/contrib/python/zstandard/py3/c-ext/bufferutil.c
new file mode 100644
index 00000000000..3d77e331e75
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/bufferutil.c
@@ -0,0 +1,578 @@
+/**
+ * Copyright (c) 2017-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void BufferWithSegments_dealloc(ZstdBufferWithSegments *self) {
+    /* Backing memory is either canonically owned by a Py_buffer or by us. */
+    if (self->parent.buf) {
+        PyBuffer_Release(&self->parent);
+    }
+    else if (self->useFree) {
+        free(self->data);
+    }
+    else {
+        PyMem_Free(self->data);
+    }
+
+    self->data = NULL;
+
+    if (self->useFree) {
+        free(self->segments);
+    }
+    else {
+        PyMem_Free(self->segments);
+    }
+
+    self->segments = NULL;
+
+    PyObject_Del(self);
+}
+
+static int BufferWithSegments_init(ZstdBufferWithSegments *self, PyObject *args,
+                                   PyObject *kwargs) {
+    static char *kwlist[] = {"data", "segments", NULL};
+
+    Py_buffer segments;
+    Py_ssize_t segmentCount;
+    Py_ssize_t i;
+
+    memset(&self->parent, 0, sizeof(self->parent));
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*y*:BufferWithSegments",
+                                     kwlist, &self->parent, &segments)) {
+        return -1;
+    }
+
+    if (segments.len % sizeof(BufferSegment)) {
+        PyErr_Format(PyExc_ValueError,
+                     "segments array size is not a multiple of %zu",
+                     sizeof(BufferSegment));
+        goto except;
+    }
+
+    segmentCount = segments.len / sizeof(BufferSegment);
+
+    /* Validate segments data, as blindly trusting it could lead to arbitrary
+    memory access. */
+    for (i = 0; i < segmentCount; i++) {
+        BufferSegment *segment = &((BufferSegment *)(segments.buf))[i];
+
+        if (segment->offset + segment->length >
+            (unsigned long long)self->parent.len) {
+            PyErr_SetString(PyExc_ValueError,
+                            "offset within segments array references memory "
+                            "outside buffer");
+            goto except;
+            return -1;
+        }
+    }
+
+    /* Make a copy of the segments data. It is cheap to do so and is a guard
+       against caller changing offsets, which has security implications. */
+    self->segments = PyMem_Malloc(segments.len);
+    if (!self->segments) {
+        PyErr_NoMemory();
+        goto except;
+    }
+
+    memcpy(self->segments, segments.buf, segments.len);
+    PyBuffer_Release(&segments);
+
+    self->data = self->parent.buf;
+    self->dataSize = self->parent.len;
+    self->segmentCount = segmentCount;
+
+    return 0;
+
+except:
+    PyBuffer_Release(&self->parent);
+    PyBuffer_Release(&segments);
+    return -1;
+}
+
+/**
+ * Construct a BufferWithSegments from existing memory and offsets.
+ *
+ * Ownership of the backing memory and BufferSegments will be transferred to
+ * the created object and freed when the BufferWithSegments is destroyed.
+ */
+ZstdBufferWithSegments *
+BufferWithSegments_FromMemory(void *data, unsigned long long dataSize,
+                              BufferSegment *segments,
+                              Py_ssize_t segmentsSize) {
+    ZstdBufferWithSegments *result = NULL;
+    Py_ssize_t i;
+
+    if (NULL == data) {
+        PyErr_SetString(PyExc_ValueError, "data is NULL");
+        return NULL;
+    }
+
+    if (NULL == segments) {
+        PyErr_SetString(PyExc_ValueError, "segments is NULL");
+        return NULL;
+    }
+
+    for (i = 0; i < segmentsSize; i++) {
+        BufferSegment *segment = &segments[i];
+
+        if (segment->offset + segment->length > dataSize) {
+            PyErr_SetString(PyExc_ValueError,
+                            "offset in segments overflows buffer size");
+            return NULL;
+        }
+    }
+
+    result = PyObject_New(ZstdBufferWithSegments, ZstdBufferWithSegmentsType);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    result->useFree = 0;
+
+    memset(&result->parent, 0, sizeof(result->parent));
+    result->data = data;
+    result->dataSize = dataSize;
+    result->segments = segments;
+    result->segmentCount = segmentsSize;
+
+    return result;
+}
+
+static Py_ssize_t BufferWithSegments_length(ZstdBufferWithSegments *self) {
+    return self->segmentCount;
+}
+
+static ZstdBufferSegment *BufferWithSegments_item(ZstdBufferWithSegments *self,
+                                                  Py_ssize_t i) {
+    ZstdBufferSegment *result = NULL;
+
+    if (i < 0) {
+        PyErr_SetString(PyExc_IndexError, "offset must be non-negative");
+        return NULL;
+    }
+
+    if (i >= self->segmentCount) {
+        PyErr_Format(PyExc_IndexError, "offset must be less than %zd",
+                     self->segmentCount);
+        return NULL;
+    }
+
+    if (self->segments[i].length > PY_SSIZE_T_MAX) {
+        PyErr_Format(PyExc_ValueError,
+                     "item at offset %zd is too large for this platform", i);
+        return NULL;
+    }
+
+    result = (ZstdBufferSegment *)PyObject_CallObject(
+        (PyObject *)ZstdBufferSegmentType, NULL);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    result->parent = (PyObject *)self;
+    Py_INCREF(self);
+
+    result->data = (char *)self->data + self->segments[i].offset;
+    result->dataSize = (Py_ssize_t)self->segments[i].length;
+    result->offset = self->segments[i].offset;
+
+    return result;
+}
+
+static int BufferWithSegments_getbuffer(ZstdBufferWithSegments *self,
+                                        Py_buffer *view, int flags) {
+    if (self->dataSize > PY_SSIZE_T_MAX) {
+        view->obj = NULL;
+        PyErr_SetString(PyExc_BufferError,
+                        "buffer is too large for this platform");
+        return -1;
+    }
+
+    return PyBuffer_FillInfo(view, (PyObject *)self, self->data,
+                             (Py_ssize_t)self->dataSize, 1, flags);
+}
+
+static PyObject *BufferWithSegments_tobytes(ZstdBufferWithSegments *self) {
+    if (self->dataSize > PY_SSIZE_T_MAX) {
+        PyErr_SetString(PyExc_ValueError,
+                        "buffer is too large for this platform");
+        return NULL;
+    }
+
+    return PyBytes_FromStringAndSize(self->data, (Py_ssize_t)self->dataSize);
+}
+
+static ZstdBufferSegments *
+BufferWithSegments_segments(ZstdBufferWithSegments *self) {
+    ZstdBufferSegments *result = (ZstdBufferSegments *)PyObject_CallObject(
+        (PyObject *)ZstdBufferSegmentsType, NULL);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    result->parent = (PyObject *)self;
+    Py_INCREF(self);
+    result->segments = self->segments;
+    result->segmentCount = self->segmentCount;
+
+    return result;
+}
+
+#if PY_VERSION_HEX < 0x03090000
+static PyBufferProcs BufferWithSegments_as_buffer = {
+    (getbufferproc)BufferWithSegments_getbuffer, /* bf_getbuffer */
+    0                                            /* bf_releasebuffer */
+};
+#endif
+
+static PyMethodDef BufferWithSegments_methods[] = {
+    {"segments", (PyCFunction)BufferWithSegments_segments, METH_NOARGS, NULL},
+    {"tobytes", (PyCFunction)BufferWithSegments_tobytes, METH_NOARGS, NULL},
+    {NULL, NULL}};
+
+static PyMemberDef BufferWithSegments_members[] = {
+    {"size", T_ULONGLONG, offsetof(ZstdBufferWithSegments, dataSize), READONLY,
+     "total size of the buffer in bytes"},
+    {NULL}};
+
+PyType_Slot ZstdBufferWithSegmentsSlots[] = {
+    {Py_tp_dealloc, BufferWithSegments_dealloc},
+    {Py_sq_length, BufferWithSegments_length},
+    {Py_sq_item, BufferWithSegments_item},
+#if PY_VERSION_HEX >= 0x03090000
+    {Py_bf_getbuffer, BufferWithSegments_getbuffer},
+#endif
+    {Py_tp_methods, BufferWithSegments_methods},
+    {Py_tp_members, BufferWithSegments_members},
+    {Py_tp_init, BufferWithSegments_init},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdBufferWithSegmentsSpec = {
+    "zstd.BufferWithSegments",
+    sizeof(ZstdBufferWithSegments),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    ZstdBufferWithSegmentsSlots,
+};
+
+PyTypeObject *ZstdBufferWithSegmentsType;
+
+static void BufferSegments_dealloc(ZstdBufferSegments *self) {
+    Py_CLEAR(self->parent);
+    PyObject_Del(self);
+}
+
+static int BufferSegments_getbuffer(ZstdBufferSegments *self, Py_buffer *view,
+                                    int flags) {
+    return PyBuffer_FillInfo(view, (PyObject *)self, (void *)self->segments,
+                             self->segmentCount * sizeof(BufferSegment), 1,
+                             flags);
+}
+
+PyType_Slot ZstdBufferSegmentsSlots[] = {
+    {Py_tp_dealloc, BufferSegments_dealloc},
+#if PY_VERSION_HEX >= 0x03090000
+    {Py_bf_getbuffer, BufferSegments_getbuffer},
+#endif
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdBufferSegmentsSpec = {
+    "zstd.BufferSegments",
+    sizeof(ZstdBufferSegments),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    ZstdBufferSegmentsSlots,
+};
+
+#if PY_VERSION_HEX < 0x03090000
+static PyBufferProcs BufferSegments_as_buffer = {
+    (getbufferproc)BufferSegments_getbuffer, 0};
+#endif
+
+PyTypeObject *ZstdBufferSegmentsType;
+
+static void BufferSegment_dealloc(ZstdBufferSegment *self) {
+    Py_CLEAR(self->parent);
+    PyObject_Del(self);
+}
+
+static Py_ssize_t BufferSegment_length(ZstdBufferSegment *self) {
+    return self->dataSize;
+}
+
+static int BufferSegment_getbuffer(ZstdBufferSegment *self, Py_buffer *view,
+                                   int flags) {
+    return PyBuffer_FillInfo(view, (PyObject *)self, self->data, self->dataSize,
+                             1, flags);
+}
+
+static PyObject *BufferSegment_tobytes(ZstdBufferSegment *self) {
+    return PyBytes_FromStringAndSize(self->data, self->dataSize);
+}
+
+#if PY_VERSION_HEX < 0x03090000
+static PyBufferProcs BufferSegment_as_buffer = {
+    (getbufferproc)BufferSegment_getbuffer, 0};
+#endif
+
+static PyMethodDef BufferSegment_methods[] = {
+    {"tobytes", (PyCFunction)BufferSegment_tobytes, METH_NOARGS, NULL},
+    {NULL, NULL}};
+
+static PyMemberDef BufferSegment_members[] = {
+    {"offset", T_ULONGLONG, offsetof(ZstdBufferSegment, offset), READONLY,
+     "offset of segment within parent buffer"},
+    {NULL}};
+
+PyType_Slot ZstdBufferSegmentSlots[] = {
+    {Py_tp_dealloc, BufferSegment_dealloc},
+    {Py_sq_length, BufferSegment_length},
+#if PY_VERSION_HEX >= 0x03090000
+    {Py_bf_getbuffer, BufferSegment_getbuffer},
+#endif
+    {Py_tp_methods, BufferSegment_methods},
+    {Py_tp_members, BufferSegment_members},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdBufferSegmentSpec = {
+    "zstd.BufferSegment",
+    sizeof(ZstdBufferSegment),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    ZstdBufferSegmentSlots,
+};
+
+PyTypeObject *ZstdBufferSegmentType;
+
+static void
+BufferWithSegmentsCollection_dealloc(ZstdBufferWithSegmentsCollection *self) {
+    Py_ssize_t i;
+
+    if (self->firstElements) {
+        PyMem_Free(self->firstElements);
+        self->firstElements = NULL;
+    }
+
+    if (self->buffers) {
+        for (i = 0; i < self->bufferCount; i++) {
+            Py_CLEAR(self->buffers[i]);
+        }
+
+        PyMem_Free(self->buffers);
+        self->buffers = NULL;
+    }
+
+    PyObject_Del(self);
+}
+
+static int
+BufferWithSegmentsCollection_init(ZstdBufferWithSegmentsCollection *self,
+                                  PyObject *args) {
+    Py_ssize_t size;
+    Py_ssize_t i;
+    Py_ssize_t offset = 0;
+
+    size = PyTuple_Size(args);
+    if (-1 == size) {
+        return -1;
+    }
+
+    if (0 == size) {
+        PyErr_SetString(PyExc_ValueError, "must pass at least 1 argument");
+        return -1;
+    }
+
+    for (i = 0; i < size; i++) {
+        PyObject *item = PyTuple_GET_ITEM(args, i);
+        if (!PyObject_TypeCheck(item, ZstdBufferWithSegmentsType)) {
+            PyErr_SetString(PyExc_TypeError,
+                            "arguments must be BufferWithSegments instances");
+            return -1;
+        }
+
+        if (0 == ((ZstdBufferWithSegments *)item)->segmentCount ||
+            0 == ((ZstdBufferWithSegments *)item)->dataSize) {
+            PyErr_SetString(PyExc_ValueError,
+                            "ZstdBufferWithSegments cannot be empty");
+            return -1;
+        }
+    }
+
+    self->buffers = PyMem_Malloc(size * sizeof(ZstdBufferWithSegments *));
+    if (NULL == self->buffers) {
+        PyErr_NoMemory();
+        return -1;
+    }
+
+    self->firstElements = PyMem_Malloc(size * sizeof(Py_ssize_t));
+    if (NULL == self->firstElements) {
+        PyMem_Free(self->buffers);
+        self->buffers = NULL;
+        PyErr_NoMemory();
+        return -1;
+    }
+
+    self->bufferCount = size;
+
+    for (i = 0; i < size; i++) {
+        ZstdBufferWithSegments *item =
+            (ZstdBufferWithSegments *)PyTuple_GET_ITEM(args, i);
+
+        self->buffers[i] = item;
+        Py_INCREF(item);
+
+        if (i > 0) {
+            self->firstElements[i - 1] = offset;
+        }
+
+        offset += item->segmentCount;
+    }
+
+    self->firstElements[size - 1] = offset;
+
+    return 0;
+}
+
+static PyObject *
+BufferWithSegmentsCollection_size(ZstdBufferWithSegmentsCollection *self) {
+    Py_ssize_t i;
+    Py_ssize_t j;
+    unsigned long long size = 0;
+
+    for (i = 0; i < self->bufferCount; i++) {
+        for (j = 0; j < self->buffers[i]->segmentCount; j++) {
+            size += self->buffers[i]->segments[j].length;
+        }
+    }
+
+    return PyLong_FromUnsignedLongLong(size);
+}
+
+Py_ssize_t
+BufferWithSegmentsCollection_length(ZstdBufferWithSegmentsCollection *self) {
+    return self->firstElements[self->bufferCount - 1];
+}
+
+static ZstdBufferSegment *
+BufferWithSegmentsCollection_item(ZstdBufferWithSegmentsCollection *self,
+                                  Py_ssize_t i) {
+    Py_ssize_t bufferOffset;
+
+    if (i < 0) {
+        PyErr_SetString(PyExc_IndexError, "offset must be non-negative");
+        return NULL;
+    }
+
+    if (i >= BufferWithSegmentsCollection_length(self)) {
+        PyErr_Format(PyExc_IndexError, "offset must be less than %zd",
+                     BufferWithSegmentsCollection_length(self));
+        return NULL;
+    }
+
+    for (bufferOffset = 0; bufferOffset < self->bufferCount; bufferOffset++) {
+        Py_ssize_t offset = 0;
+
+        if (i < self->firstElements[bufferOffset]) {
+            if (bufferOffset > 0) {
+                offset = self->firstElements[bufferOffset - 1];
+            }
+
+            return BufferWithSegments_item(self->buffers[bufferOffset],
+                                           i - offset);
+        }
+    }
+
+    PyErr_SetString(ZstdError,
+                    "error resolving segment; this should not happen");
+    return NULL;
+}
+
+static PyMethodDef BufferWithSegmentsCollection_methods[] = {
+    {"size", (PyCFunction)BufferWithSegmentsCollection_size, METH_NOARGS,
+     PyDoc_STR("total size in bytes of all segments")},
+    {NULL, NULL}};
+
+PyType_Slot ZstdBufferWithSegmentsCollectionSlots[] = {
+    {Py_tp_dealloc, BufferWithSegmentsCollection_dealloc},
+    {Py_sq_length, BufferWithSegmentsCollection_length},
+    {Py_sq_item, BufferWithSegmentsCollection_item},
+    {Py_tp_methods, BufferWithSegmentsCollection_methods},
+    {Py_tp_init, BufferWithSegmentsCollection_init},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdBufferWithSegmentsCollectionSpec = {
+    "zstd.BufferWithSegmentsCollection",
+    sizeof(ZstdBufferWithSegmentsCollection),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    ZstdBufferWithSegmentsCollectionSlots,
+};
+
+PyTypeObject *ZstdBufferWithSegmentsCollectionType;
+
+void bufferutil_module_init(PyObject *mod) {
+    ZstdBufferWithSegmentsType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdBufferWithSegmentsSpec);
+#if PY_VERSION_HEX < 0x03090000
+    ZstdBufferWithSegmentsType->tp_as_buffer = &BufferWithSegments_as_buffer;
+#endif
+    if (PyType_Ready(ZstdBufferWithSegmentsType) < 0) {
+        return;
+    }
+
+    Py_INCREF(ZstdBufferWithSegmentsType);
+    PyModule_AddObject(mod, "BufferWithSegments",
+                       (PyObject *)ZstdBufferWithSegmentsType);
+
+    ZstdBufferSegmentsType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdBufferSegmentsSpec);
+#if PY_VERSION_HEX < 0x03090000
+    ZstdBufferSegmentsType->tp_as_buffer = &BufferSegments_as_buffer;
+#endif
+    if (PyType_Ready(ZstdBufferSegmentsType) < 0) {
+        return;
+    }
+
+    Py_INCREF(ZstdBufferSegmentsType);
+    PyModule_AddObject(mod, "BufferSegments",
+                       (PyObject *)ZstdBufferSegmentsType);
+
+    ZstdBufferSegmentType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdBufferSegmentSpec);
+#if PY_VERSION_HEX < 0x03090000
+    ZstdBufferSegmentType->tp_as_buffer = &BufferSegment_as_buffer;
+#endif
+    if (PyType_Ready(ZstdBufferSegmentType) < 0) {
+        return;
+    }
+
+    Py_INCREF(ZstdBufferSegmentType);
+    PyModule_AddObject(mod, "BufferSegment", (PyObject *)ZstdBufferSegmentType);
+
+    ZstdBufferWithSegmentsCollectionType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdBufferWithSegmentsCollectionSpec);
+    if (PyType_Ready(ZstdBufferWithSegmentsCollectionType) < 0) {
+        return;
+    }
+
+    Py_INCREF(ZstdBufferWithSegmentsCollectionType);
+    PyModule_AddObject(mod, "BufferWithSegmentsCollection",
+                       (PyObject *)ZstdBufferWithSegmentsCollectionType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressionchunker.c b/contrib/python/zstandard/py3/c-ext/compressionchunker.c
new file mode 100644
index 00000000000..bdf6f5d8ee5
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressionchunker.c
@@ -0,0 +1,310 @@
+/**
+ * Copyright (c) 2018-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void
+ZstdCompressionChunkerIterator_dealloc(ZstdCompressionChunkerIterator *self) {
+    Py_XDECREF(self->chunker);
+
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdCompressionChunkerIterator_iter(PyObject *self) {
+    Py_INCREF(self);
+    return self;
+}
+
+static PyObject *
+ZstdCompressionChunkerIterator_iternext(ZstdCompressionChunkerIterator *self) {
+    size_t zresult;
+    PyObject *chunk;
+    ZstdCompressionChunker *chunker = self->chunker;
+    ZSTD_EndDirective zFlushMode;
+
+    if (self->mode != compressionchunker_mode_normal &&
+        chunker->input.pos != chunker->input.size) {
+        PyErr_SetString(ZstdError, "input should have been fully consumed "
+                                   "before calling flush() or finish()");
+        return NULL;
+    }
+
+    if (chunker->finished) {
+        return NULL;
+    }
+
+    /* If we have data left in the input, consume it. */
+    while (chunker->input.pos < chunker->input.size) {
+        Py_BEGIN_ALLOW_THREADS zresult =
+            ZSTD_compressStream2(chunker->compressor->cctx, &chunker->output,
+                                 &chunker->input, ZSTD_e_continue);
+        Py_END_ALLOW_THREADS
+
+            /* Input is fully consumed. */
+            if (chunker->input.pos == chunker->input.size) {
+            chunker->input.src = NULL;
+            chunker->input.pos = 0;
+            chunker->input.size = 0;
+            PyBuffer_Release(&chunker->inBuffer);
+        }
+
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd compress error: %s",
+                         ZSTD_getErrorName(zresult));
+            return NULL;
+        }
+
+        /* If it produced a full output chunk, emit it. */
+        if (chunker->output.pos == chunker->output.size) {
+            chunk = PyBytes_FromStringAndSize(chunker->output.dst,
+                                              chunker->output.pos);
+            if (!chunk) {
+                return NULL;
+            }
+
+            chunker->output.pos = 0;
+
+            return chunk;
+        }
+
+        /* Else continue to compress available input data. */
+    }
+
+    /* We also need this here for the special case of an empty input buffer. */
+    if (chunker->input.pos == chunker->input.size) {
+        chunker->input.src = NULL;
+        chunker->input.pos = 0;
+        chunker->input.size = 0;
+        PyBuffer_Release(&chunker->inBuffer);
+    }
+
+    /* No more input data. A partial chunk may be in chunker->output.
+     * If we're in normal compression mode, we're done. Otherwise if we're in
+     * flush or finish mode, we need to emit what data remains.
+     */
+    if (self->mode == compressionchunker_mode_normal) {
+        /* We don't need to set StopIteration. */
+        return NULL;
+    }
+
+    if (self->mode == compressionchunker_mode_flush) {
+        zFlushMode = ZSTD_e_flush;
+    }
+    else if (self->mode == compressionchunker_mode_finish) {
+        zFlushMode = ZSTD_e_end;
+    }
+    else {
+        PyErr_SetString(ZstdError,
+                        "unhandled compression mode; this should never happen");
+        return NULL;
+    }
+
+    Py_BEGIN_ALLOW_THREADS zresult =
+        ZSTD_compressStream2(chunker->compressor->cctx, &chunker->output,
+                             &chunker->input, zFlushMode);
+    Py_END_ALLOW_THREADS
+
+        if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "zstd compress error: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    if (!zresult && chunker->output.pos == 0) {
+        return NULL;
+    }
+
+    chunk = PyBytes_FromStringAndSize(chunker->output.dst, chunker->output.pos);
+    if (!chunk) {
+        return NULL;
+    }
+
+    chunker->output.pos = 0;
+
+    if (!zresult && self->mode == compressionchunker_mode_finish) {
+        chunker->finished = 1;
+    }
+
+    return chunk;
+}
+
+PyType_Slot ZstdCompressionChunkerIteratorSlots[] = {
+    {Py_tp_dealloc, ZstdCompressionChunkerIterator_dealloc},
+    {Py_tp_iter, ZstdCompressionChunkerIterator_iter},
+    {Py_tp_iternext, ZstdCompressionChunkerIterator_iternext},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionChunkerIteratorSpec = {
+    "zstd.ZstdCompressionChunkerIterator",
+    sizeof(ZstdCompressionChunkerIterator),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressionChunkerIteratorSlots,
+};
+
+PyTypeObject *ZstdCompressionChunkerIteratorType;
+
+static void ZstdCompressionChunker_dealloc(ZstdCompressionChunker *self) {
+    PyBuffer_Release(&self->inBuffer);
+    self->input.src = NULL;
+
+    PyMem_Free(self->output.dst);
+    self->output.dst = NULL;
+
+    Py_XDECREF(self->compressor);
+
+    PyObject_Del(self);
+}
+
+static ZstdCompressionChunkerIterator *
+ZstdCompressionChunker_compress(ZstdCompressionChunker *self, PyObject *args,
+                                PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    ZstdCompressionChunkerIterator *result;
+
+    if (self->finished) {
+        PyErr_SetString(ZstdError,
+                        "cannot call compress() after compression finished");
+        return NULL;
+    }
+
+    if (self->inBuffer.obj) {
+        PyErr_SetString(ZstdError, "cannot perform operation before consuming "
+                                   "output from previous operation");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:compress", kwlist,
+                                     &self->inBuffer)) {
+        return NULL;
+    }
+
+    result = (ZstdCompressionChunkerIterator *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionChunkerIteratorType, NULL);
+    if (!result) {
+        PyBuffer_Release(&self->inBuffer);
+        return NULL;
+    }
+
+    self->input.src = self->inBuffer.buf;
+    self->input.size = self->inBuffer.len;
+    self->input.pos = 0;
+
+    result->chunker = self;
+    Py_INCREF(result->chunker);
+
+    result->mode = compressionchunker_mode_normal;
+
+    return result;
+}
+
+static ZstdCompressionChunkerIterator *
+ZstdCompressionChunker_finish(ZstdCompressionChunker *self) {
+    ZstdCompressionChunkerIterator *result;
+
+    if (self->finished) {
+        PyErr_SetString(ZstdError,
+                        "cannot call finish() after compression finished");
+        return NULL;
+    }
+
+    if (self->inBuffer.obj) {
+        PyErr_SetString(ZstdError, "cannot call finish() before consuming "
+                                   "output from previous operation");
+        return NULL;
+    }
+
+    result = (ZstdCompressionChunkerIterator *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionChunkerIteratorType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    result->chunker = self;
+    Py_INCREF(result->chunker);
+
+    result->mode = compressionchunker_mode_finish;
+
+    return result;
+}
+
+static ZstdCompressionChunkerIterator *
+ZstdCompressionChunker_flush(ZstdCompressionChunker *self, PyObject *args,
+                             PyObject *kwargs) {
+    ZstdCompressionChunkerIterator *result;
+
+    if (self->finished) {
+        PyErr_SetString(ZstdError,
+                        "cannot call flush() after compression finished");
+        return NULL;
+    }
+
+    if (self->inBuffer.obj) {
+        PyErr_SetString(ZstdError, "cannot call flush() before consuming "
+                                   "output from previous operation");
+        return NULL;
+    }
+
+    result = (ZstdCompressionChunkerIterator *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionChunkerIteratorType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    result->chunker = self;
+    Py_INCREF(result->chunker);
+
+    result->mode = compressionchunker_mode_flush;
+
+    return result;
+}
+
+static PyMethodDef ZstdCompressionChunker_methods[] = {
+    {"compress", (PyCFunction)ZstdCompressionChunker_compress,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("compress data")},
+    {"finish", (PyCFunction)ZstdCompressionChunker_finish, METH_NOARGS,
+     PyDoc_STR("finish compression operation")},
+    {"flush", (PyCFunction)ZstdCompressionChunker_flush,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("finish compression operation")},
+    {NULL, NULL}};
+
+PyType_Slot ZstdCompressionChunkerSlots[] = {
+    {Py_tp_dealloc, ZstdCompressionChunker_dealloc},
+    {Py_tp_methods, ZstdCompressionChunker_methods},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionChunkerSpec = {
+    "zstd.ZstdCompressionChunkerType",
+    sizeof(ZstdCompressionChunker),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressionChunkerSlots,
+};
+
+PyTypeObject *ZstdCompressionChunkerType;
+
+void compressionchunker_module_init(PyObject *module) {
+    ZstdCompressionChunkerIteratorType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionChunkerIteratorSpec);
+    if (PyType_Ready(ZstdCompressionChunkerIteratorType) < 0) {
+        return;
+    }
+
+    ZstdCompressionChunkerType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionChunkerSpec);
+    if (PyType_Ready(ZstdCompressionChunkerType) < 0) {
+        return;
+    }
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressiondict.c b/contrib/python/zstandard/py3/c-ext/compressiondict.c
new file mode 100644
index 00000000000..437601202b8
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressiondict.c
@@ -0,0 +1,348 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+ZstdCompressionDict *train_dictionary(PyObject *self, PyObject *args,
+                                      PyObject *kwargs) {
+    static char *kwlist[] = {
+        "dict_size", "samples",     "k",     "d",
+        "f",         "split_point", "accel", "notifications",
+        "dict_id",   "level",       "steps", "threads",
+        NULL};
+
+    size_t capacity;
+    PyObject *samples;
+    unsigned k = 0;
+    unsigned d = 0;
+    unsigned f = 0;
+    double splitPoint = 0.0;
+    unsigned accel = 0;
+    unsigned notifications = 0;
+    unsigned dictID = 0;
+    int level = 0;
+    unsigned steps = 0;
+    int threads = 0;
+    ZDICT_fastCover_params_t params;
+    Py_ssize_t samplesLen;
+    Py_ssize_t i;
+    size_t samplesSize = 0;
+    void *sampleBuffer = NULL;
+    size_t *sampleSizes = NULL;
+    void *sampleOffset;
+    Py_ssize_t sampleSize;
+    void *dict = NULL;
+    size_t zresult;
+    ZstdCompressionDict *result = NULL;
+
+    if (!PyArg_ParseTupleAndKeywords(
+            args, kwargs, "nO!|IIIdIIIiIi:train_dictionary", kwlist, &capacity,
+            &PyList_Type, &samples, &k, &d, &f, &splitPoint, &accel,
+            &notifications, &dictID, &level, &steps, &threads)) {
+        return NULL;
+    }
+
+    if (threads < 0) {
+        threads = cpu_count();
+    }
+
+    if (!steps && !threads) {
+        /* Defaults from ZDICT_trainFromBuffer() */
+        d = d ? d : 8;
+        steps = steps ? steps : 4;
+        level = level ? level : 3;
+    }
+
+    memset(&params, 0, sizeof(params));
+    params.k = k;
+    params.d = d;
+    params.f = f;
+    params.steps = steps;
+    params.nbThreads = threads;
+    params.splitPoint = splitPoint;
+    params.accel = accel;
+
+    params.zParams.compressionLevel = level;
+    params.zParams.dictID = dictID;
+    params.zParams.notificationLevel = notifications;
+
+    /* Figure out total size of input samples. */
+    samplesLen = PyList_Size(samples);
+    for (i = 0; i < samplesLen; i++) {
+        PyObject *sampleItem = PyList_GET_ITEM(samples, i);
+
+        if (!PyBytes_Check(sampleItem)) {
+            PyErr_SetString(PyExc_ValueError, "samples must be bytes");
+            return NULL;
+        }
+        samplesSize += PyBytes_GET_SIZE(sampleItem);
+    }
+
+    sampleBuffer = PyMem_Malloc(samplesSize);
+    if (!sampleBuffer) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+
+    sampleSizes = PyMem_Malloc(samplesLen * sizeof(size_t));
+    if (!sampleSizes) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+
+    sampleOffset = sampleBuffer;
+    for (i = 0; i < samplesLen; i++) {
+        PyObject *sampleItem = PyList_GET_ITEM(samples, i);
+        sampleSize = PyBytes_GET_SIZE(sampleItem);
+        sampleSizes[i] = sampleSize;
+        memcpy(sampleOffset, PyBytes_AS_STRING(sampleItem), sampleSize);
+        sampleOffset = (char *)sampleOffset + sampleSize;
+    }
+
+    dict = PyMem_Malloc(capacity);
+    if (!dict) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+
+    Py_BEGIN_ALLOW_THREADS zresult = ZDICT_optimizeTrainFromBuffer_fastCover(
+        dict, capacity, sampleBuffer, sampleSizes, (unsigned)samplesLen,
+        &params);
+    Py_END_ALLOW_THREADS
+
+        if (ZDICT_isError(zresult)) {
+        PyMem_Free(dict);
+        PyErr_Format(ZstdError, "cannot train dict: %s",
+                     ZDICT_getErrorName(zresult));
+        goto finally;
+    }
+
+    result = PyObject_New(ZstdCompressionDict, ZstdCompressionDictType);
+    if (!result) {
+        PyMem_Free(dict);
+        goto finally;
+    }
+
+    result->dictData = dict;
+    result->dictSize = zresult;
+    result->dictType = ZSTD_dct_fullDict;
+    result->d = params.d;
+    result->k = params.k;
+    result->cdict = NULL;
+    result->ddict = NULL;
+
+finally:
+    PyMem_Free(sampleBuffer);
+    PyMem_Free(sampleSizes);
+
+    return result;
+}
+
+int ensure_ddict(ZstdCompressionDict *dict) {
+    if (dict->ddict) {
+        return 0;
+    }
+
+    Py_BEGIN_ALLOW_THREADS dict->ddict = ZSTD_createDDict_advanced(
+        dict->dictData, dict->dictSize, ZSTD_dlm_byRef, dict->dictType,
+        ZSTD_defaultCMem);
+    Py_END_ALLOW_THREADS if (!dict->ddict) {
+        PyErr_SetString(ZstdError, "could not create decompression dict");
+        return 1;
+    }
+
+    return 0;
+}
+
+static int ZstdCompressionDict_init(ZstdCompressionDict *self, PyObject *args,
+                                    PyObject *kwargs) {
+    static char *kwlist[] = {"data", "dict_type", NULL};
+
+    int result = -1;
+    Py_buffer source;
+    unsigned dictType = ZSTD_dct_auto;
+
+    self->dictData = NULL;
+    self->dictSize = 0;
+    self->cdict = NULL;
+    self->ddict = NULL;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|I:ZstdCompressionDict",
+                                     kwlist, &source, &dictType)) {
+        return -1;
+    }
+
+    if (dictType != ZSTD_dct_auto && dictType != ZSTD_dct_rawContent &&
+        dictType != ZSTD_dct_fullDict) {
+        PyErr_Format(
+            PyExc_ValueError,
+            "invalid dictionary load mode: %d; must use DICT_TYPE_* constants",
+            dictType);
+        goto finally;
+    }
+
+    self->dictType = dictType;
+
+    self->dictData = PyMem_Malloc(source.len);
+    if (!self->dictData) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+
+    memcpy(self->dictData, source.buf, source.len);
+    self->dictSize = source.len;
+
+    result = 0;
+
+finally:
+    PyBuffer_Release(&source);
+    return result;
+}
+
+static void ZstdCompressionDict_dealloc(ZstdCompressionDict *self) {
+    if (self->cdict) {
+        ZSTD_freeCDict(self->cdict);
+        self->cdict = NULL;
+    }
+
+    if (self->ddict) {
+        ZSTD_freeDDict(self->ddict);
+        self->ddict = NULL;
+    }
+
+    if (self->dictData) {
+        PyMem_Free(self->dictData);
+        self->dictData = NULL;
+    }
+
+    PyObject_Del(self);
+}
+
+static PyObject *
+ZstdCompressionDict_precompute_compress(ZstdCompressionDict *self,
+                                        PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"level", "compression_params", NULL};
+
+    int level = 0;
+    ZstdCompressionParametersObject *compressionParams = NULL;
+    ZSTD_compressionParameters cParams;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(
+            args, kwargs, "|iO!:precompute_compress", kwlist, &level,
+            ZstdCompressionParametersType, &compressionParams)) {
+        return NULL;
+    }
+
+    if (level && compressionParams) {
+        PyErr_SetString(PyExc_ValueError,
+                        "must only specify one of level or compression_params");
+        return NULL;
+    }
+
+    if (!level && !compressionParams) {
+        PyErr_SetString(PyExc_ValueError,
+                        "must specify one of level or compression_params");
+        return NULL;
+    }
+
+    if (self->cdict) {
+        zresult = ZSTD_freeCDict(self->cdict);
+        self->cdict = NULL;
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "unable to free CDict: %s",
+                         ZSTD_getErrorName(zresult));
+            return NULL;
+        }
+    }
+
+    if (level) {
+        cParams = ZSTD_getCParams(level, 0, self->dictSize);
+    }
+    else {
+        if (to_cparams(compressionParams, &cParams)) {
+            return NULL;
+        }
+    }
+
+    assert(!self->cdict);
+    self->cdict = ZSTD_createCDict_advanced(self->dictData, self->dictSize,
+                                            ZSTD_dlm_byRef, self->dictType,
+                                            cParams, ZSTD_defaultCMem);
+
+    if (!self->cdict) {
+        PyErr_SetString(ZstdError, "unable to precompute dictionary");
+        return NULL;
+    }
+
+    Py_RETURN_NONE;
+}
+
+static PyObject *ZstdCompressionDict_dict_id(ZstdCompressionDict *self) {
+    unsigned dictID = ZDICT_getDictID(self->dictData, self->dictSize);
+
+    return PyLong_FromLong(dictID);
+}
+
+static PyObject *ZstdCompressionDict_as_bytes(ZstdCompressionDict *self) {
+    return PyBytes_FromStringAndSize(self->dictData, self->dictSize);
+}
+
+static PyMethodDef ZstdCompressionDict_methods[] = {
+    {"dict_id", (PyCFunction)ZstdCompressionDict_dict_id, METH_NOARGS,
+     PyDoc_STR("dict_id() -- obtain the numeric dictionary ID")},
+    {"as_bytes", (PyCFunction)ZstdCompressionDict_as_bytes, METH_NOARGS,
+     PyDoc_STR("as_bytes() -- obtain the raw bytes constituting the dictionary "
+               "data")},
+    {"precompute_compress",
+     (PyCFunction)ZstdCompressionDict_precompute_compress,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {NULL, NULL}};
+
+static PyMemberDef ZstdCompressionDict_members[] = {
+    {"k", T_UINT, offsetof(ZstdCompressionDict, k), READONLY, "segment size"},
+    {"d", T_UINT, offsetof(ZstdCompressionDict, d), READONLY, "dmer size"},
+    {NULL}};
+
+static Py_ssize_t ZstdCompressionDict_length(ZstdCompressionDict *self) {
+    return self->dictSize;
+}
+
+PyType_Slot ZstdCompressionDictSlots[] = {
+    {Py_tp_dealloc, ZstdCompressionDict_dealloc},
+    {Py_sq_length, ZstdCompressionDict_length},
+    {Py_tp_methods, ZstdCompressionDict_methods},
+    {Py_tp_members, ZstdCompressionDict_members},
+    {Py_tp_init, ZstdCompressionDict_init},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionDictSpec = {
+    "zstd.ZstdCompressionDict",
+    sizeof(ZstdCompressionDict),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressionDictSlots,
+};
+
+PyTypeObject *ZstdCompressionDictType;
+
+void compressiondict_module_init(PyObject *mod) {
+    ZstdCompressionDictType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionDictSpec);
+    if (PyType_Ready(ZstdCompressionDictType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdCompressionDictType);
+    PyModule_AddObject(mod, "ZstdCompressionDict",
+                       (PyObject *)ZstdCompressionDictType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressionparams.c b/contrib/python/zstandard/py3/c-ext/compressionparams.c
new file mode 100644
index 00000000000..0cc7cbbe6e4
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressionparams.c
@@ -0,0 +1,501 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+int set_parameter(ZSTD_CCtx_params *params, ZSTD_cParameter param, int value) {
+    size_t zresult = ZSTD_CCtxParams_setParameter(params, param, value);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError,
+                     "unable to set compression context parameter: %s",
+                     ZSTD_getErrorName(zresult));
+        return 1;
+    }
+
+    return 0;
+}
+
+#define TRY_SET_PARAMETER(params, param, value)                                \
+    if (set_parameter(params, param, value))                                   \
+        return -1;
+
+#define TRY_COPY_PARAMETER(source, dest, param)                                \
+    {                                                                          \
+        int result;                                                            \
+        size_t zresult = ZSTD_CCtxParams_getParameter(source, param, &result); \
+        if (ZSTD_isError(zresult)) {                                           \
+            return 1;                                                          \
+        }                                                                      \
+        zresult = ZSTD_CCtxParams_setParameter(dest, param, result);           \
+        if (ZSTD_isError(zresult)) {                                           \
+            return 1;                                                          \
+        }                                                                      \
+    }
+
+int set_parameters(ZSTD_CCtx_params *params,
+                   ZstdCompressionParametersObject *obj) {
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_nbWorkers);
+
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_format);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_compressionLevel);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_windowLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_hashLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_chainLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_searchLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_minMatch);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_targetLength);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_strategy);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_contentSizeFlag);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_checksumFlag);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_dictIDFlag);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_jobSize);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_overlapLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_forceMaxWindow);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_enableLongDistanceMatching);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmHashLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmMinMatch);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmBucketSizeLog);
+    TRY_COPY_PARAMETER(obj->params, params, ZSTD_c_ldmHashRateLog);
+
+    return 0;
+}
+
+int reset_params(ZstdCompressionParametersObject *params) {
+    if (params->params) {
+        ZSTD_CCtxParams_reset(params->params);
+    }
+    else {
+        params->params = ZSTD_createCCtxParams();
+        if (!params->params) {
+            PyErr_NoMemory();
+            return 1;
+        }
+    }
+
+    return set_parameters(params->params, params);
+}
+
+#define TRY_GET_PARAMETER(params, param, value)                                \
+    {                                                                          \
+        size_t zresult = ZSTD_CCtxParams_getParameter(params, param, value);   \
+        if (ZSTD_isError(zresult)) {                                           \
+            PyErr_Format(ZstdError, "unable to retrieve parameter: %s",        \
+                         ZSTD_getErrorName(zresult));                          \
+            return 1;                                                          \
+        }                                                                      \
+    }
+
+int to_cparams(ZstdCompressionParametersObject *params,
+               ZSTD_compressionParameters *cparams) {
+    int value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_windowLog, &value);
+    cparams->windowLog = value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_chainLog, &value);
+    cparams->chainLog = value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_hashLog, &value);
+    cparams->hashLog = value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_searchLog, &value);
+    cparams->searchLog = value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_minMatch, &value);
+    cparams->minMatch = value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_targetLength, &value);
+    cparams->targetLength = value;
+
+    TRY_GET_PARAMETER(params->params, ZSTD_c_strategy, &value);
+    cparams->strategy = value;
+
+    return 0;
+}
+
+static int ZstdCompressionParameters_init(ZstdCompressionParametersObject *self,
+                                          PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"format",
+                             "compression_level",
+                             "window_log",
+                             "hash_log",
+                             "chain_log",
+                             "search_log",
+                             "min_match",
+                             "target_length",
+                             "strategy",
+                             "write_content_size",
+                             "write_checksum",
+                             "write_dict_id",
+                             "job_size",
+                             "overlap_log",
+                             "force_max_window",
+                             "enable_ldm",
+                             "ldm_hash_log",
+                             "ldm_min_match",
+                             "ldm_bucket_size_log",
+                             "ldm_hash_rate_log",
+                             "threads",
+                             NULL};
+
+    int format = 0;
+    int compressionLevel = 0;
+    int windowLog = 0;
+    int hashLog = 0;
+    int chainLog = 0;
+    int searchLog = 0;
+    int minMatch = 0;
+    int targetLength = 0;
+    int strategy = -1;
+    int contentSizeFlag = 1;
+    int checksumFlag = 0;
+    int dictIDFlag = 0;
+    int jobSize = 0;
+    int overlapLog = -1;
+    int forceMaxWindow = 0;
+    int enableLDM = 0;
+    int ldmHashLog = 0;
+    int ldmMinMatch = 0;
+    int ldmBucketSizeLog = 0;
+    int ldmHashRateLog = -1;
+    int threads = 0;
+
+    if (!PyArg_ParseTupleAndKeywords(
+            args, kwargs, "|iiiiiiiiiiiiiiiiiiiii:ZstdCompressionParameters",
+            kwlist, &format, &compressionLevel, &windowLog, &hashLog, &chainLog,
+            &searchLog, &minMatch, &targetLength, &strategy, &contentSizeFlag,
+            &checksumFlag, &dictIDFlag, &jobSize, &overlapLog, &forceMaxWindow,
+            &enableLDM, &ldmHashLog, &ldmMinMatch, &ldmBucketSizeLog,
+            &ldmHashRateLog, &threads)) {
+        return -1;
+    }
+
+    if (reset_params(self)) {
+        return -1;
+    }
+
+    if (threads < 0) {
+        threads = cpu_count();
+    }
+
+    /* We need to set ZSTD_c_nbWorkers before ZSTD_c_jobSize and
+     * ZSTD_c_overlapLog because setting ZSTD_c_nbWorkers resets the other
+     * parameters. */
+    TRY_SET_PARAMETER(self->params, ZSTD_c_nbWorkers, threads);
+
+    TRY_SET_PARAMETER(self->params, ZSTD_c_format, format);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_compressionLevel, compressionLevel);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_windowLog, windowLog);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_hashLog, hashLog);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_chainLog, chainLog);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_searchLog, searchLog);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_minMatch, minMatch);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_targetLength, targetLength);
+
+    if (strategy == -1) {
+        strategy = 0;
+    }
+
+    TRY_SET_PARAMETER(self->params, ZSTD_c_strategy, strategy);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_contentSizeFlag, contentSizeFlag);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_checksumFlag, checksumFlag);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_dictIDFlag, dictIDFlag);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_jobSize, jobSize);
+
+    if (overlapLog == -1) {
+        overlapLog = 0;
+    }
+
+    TRY_SET_PARAMETER(self->params, ZSTD_c_overlapLog, overlapLog);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_forceMaxWindow, forceMaxWindow);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_enableLongDistanceMatching,
+                      enableLDM);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_ldmHashLog, ldmHashLog);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_ldmMinMatch, ldmMinMatch);
+    TRY_SET_PARAMETER(self->params, ZSTD_c_ldmBucketSizeLog, ldmBucketSizeLog);
+
+    if (ldmHashRateLog == -1) {
+        ldmHashRateLog = 0;
+    }
+
+    TRY_SET_PARAMETER(self->params, ZSTD_c_ldmHashRateLog, ldmHashRateLog);
+
+    return 0;
+}
+
+ZstdCompressionParametersObject *
+CompressionParameters_from_level(PyObject *undef, PyObject *args,
+                                 PyObject *kwargs) {
+    int managedKwargs = 0;
+    int level;
+    PyObject *sourceSize = NULL;
+    PyObject *dictSize = NULL;
+    unsigned PY_LONG_LONG iSourceSize = 0;
+    Py_ssize_t iDictSize = 0;
+    PyObject *val;
+    ZSTD_compressionParameters params;
+    ZstdCompressionParametersObject *result = NULL;
+    int res;
+
+    if (!PyArg_ParseTuple(args, "i:from_level", &level)) {
+        return NULL;
+    }
+
+    if (!kwargs) {
+        kwargs = PyDict_New();
+        if (!kwargs) {
+            return NULL;
+        }
+        managedKwargs = 1;
+    }
+
+    sourceSize = PyDict_GetItemString(kwargs, "source_size");
+    if (sourceSize) {
+        iSourceSize = PyLong_AsUnsignedLongLong(sourceSize);
+        if (iSourceSize == (unsigned PY_LONG_LONG)(-1)) {
+            goto cleanup;
+        }
+
+        PyDict_DelItemString(kwargs, "source_size");
+    }
+
+    dictSize = PyDict_GetItemString(kwargs, "dict_size");
+    if (dictSize) {
+        iDictSize = PyLong_AsSsize_t(dictSize);
+        if (iDictSize == -1) {
+            goto cleanup;
+        }
+
+        PyDict_DelItemString(kwargs, "dict_size");
+    }
+
+    params = ZSTD_getCParams(level, iSourceSize, iDictSize);
+
+    /* Values derived from the input level and sizes are passed along to the
+       constructor. But only if a value doesn't already exist. */
+    val = PyDict_GetItemString(kwargs, "window_log");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.windowLog);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "window_log", val);
+        Py_DECREF(val);
+    }
+
+    val = PyDict_GetItemString(kwargs, "chain_log");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.chainLog);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "chain_log", val);
+        Py_DECREF(val);
+    }
+
+    val = PyDict_GetItemString(kwargs, "hash_log");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.hashLog);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "hash_log", val);
+        Py_DECREF(val);
+    }
+
+    val = PyDict_GetItemString(kwargs, "search_log");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.searchLog);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "search_log", val);
+        Py_DECREF(val);
+    }
+
+    val = PyDict_GetItemString(kwargs, "min_match");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.minMatch);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "min_match", val);
+        Py_DECREF(val);
+    }
+
+    val = PyDict_GetItemString(kwargs, "target_length");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.targetLength);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "target_length", val);
+        Py_DECREF(val);
+    }
+
+    val = PyDict_GetItemString(kwargs, "strategy");
+    if (!val) {
+        val = PyLong_FromUnsignedLong(params.strategy);
+        if (!val) {
+            goto cleanup;
+        }
+        PyDict_SetItemString(kwargs, "strategy", val);
+        Py_DECREF(val);
+    }
+
+    result = PyObject_New(ZstdCompressionParametersObject,
+                          ZstdCompressionParametersType);
+    if (!result) {
+        goto cleanup;
+    }
+
+    result->params = NULL;
+
+    val = PyTuple_New(0);
+    if (!val) {
+        Py_CLEAR(result);
+        goto cleanup;
+    }
+
+    res = ZstdCompressionParameters_init(result, val, kwargs);
+    Py_DECREF(val);
+
+    if (res) {
+        Py_CLEAR(result);
+        goto cleanup;
+    }
+
+cleanup:
+    if (managedKwargs) {
+        Py_DECREF(kwargs);
+    }
+
+    return result;
+}
+
+PyObject *ZstdCompressionParameters_estimated_compression_context_size(
+    ZstdCompressionParametersObject *self) {
+    return PyLong_FromSize_t(
+        ZSTD_estimateCCtxSize_usingCCtxParams(self->params));
+}
+
+static void
+ZstdCompressionParameters_dealloc(ZstdCompressionParametersObject *self) {
+    if (self->params) {
+        ZSTD_freeCCtxParams(self->params);
+        self->params = NULL;
+    }
+
+    PyObject_Del(self);
+}
+
+#define PARAM_GETTER(name, param)                                              \
+    PyObject *ZstdCompressionParameters_get_##name(PyObject *self,             \
+                                                   void *unused) {             \
+        int result;                                                            \
+        size_t zresult;                                                        \
+        ZstdCompressionParametersObject *p =                                   \
+            (ZstdCompressionParametersObject *)(self);                         \
+        zresult = ZSTD_CCtxParams_getParameter(p->params, param, &result);     \
+        if (ZSTD_isError(zresult)) {                                           \
+            PyErr_Format(ZstdError, "unable to get compression parameter: %s", \
+                         ZSTD_getErrorName(zresult));                          \
+            return NULL;                                                       \
+        }                                                                      \
+        return PyLong_FromLong(result);                                        \
+    }
+
+PARAM_GETTER(format, ZSTD_c_format)
+PARAM_GETTER(compression_level, ZSTD_c_compressionLevel)
+PARAM_GETTER(window_log, ZSTD_c_windowLog)
+PARAM_GETTER(hash_log, ZSTD_c_hashLog)
+PARAM_GETTER(chain_log, ZSTD_c_chainLog)
+PARAM_GETTER(search_log, ZSTD_c_searchLog)
+PARAM_GETTER(min_match, ZSTD_c_minMatch)
+PARAM_GETTER(target_length, ZSTD_c_targetLength)
+PARAM_GETTER(strategy, ZSTD_c_strategy)
+PARAM_GETTER(write_content_size, ZSTD_c_contentSizeFlag)
+PARAM_GETTER(write_checksum, ZSTD_c_checksumFlag)
+PARAM_GETTER(write_dict_id, ZSTD_c_dictIDFlag)
+PARAM_GETTER(job_size, ZSTD_c_jobSize)
+PARAM_GETTER(overlap_log, ZSTD_c_overlapLog)
+PARAM_GETTER(force_max_window, ZSTD_c_forceMaxWindow)
+PARAM_GETTER(enable_ldm, ZSTD_c_enableLongDistanceMatching)
+PARAM_GETTER(ldm_hash_log, ZSTD_c_ldmHashLog)
+PARAM_GETTER(ldm_min_match, ZSTD_c_ldmMinMatch)
+PARAM_GETTER(ldm_bucket_size_log, ZSTD_c_ldmBucketSizeLog)
+PARAM_GETTER(ldm_hash_rate_log, ZSTD_c_ldmHashRateLog)
+PARAM_GETTER(threads, ZSTD_c_nbWorkers)
+
+static PyMethodDef ZstdCompressionParameters_methods[] = {
+    {"from_level", (PyCFunction)CompressionParameters_from_level,
+     METH_VARARGS | METH_KEYWORDS | METH_STATIC, NULL},
+    {"estimated_compression_context_size",
+     (PyCFunction)ZstdCompressionParameters_estimated_compression_context_size,
+     METH_NOARGS, NULL},
+    {NULL, NULL}};
+
+#define GET_SET_ENTRY(name)                                                    \
+    { #name, ZstdCompressionParameters_get_##name, NULL, NULL, NULL }
+
+static PyGetSetDef ZstdCompressionParameters_getset[] = {
+    GET_SET_ENTRY(format),
+    GET_SET_ENTRY(compression_level),
+    GET_SET_ENTRY(window_log),
+    GET_SET_ENTRY(hash_log),
+    GET_SET_ENTRY(chain_log),
+    GET_SET_ENTRY(search_log),
+    GET_SET_ENTRY(min_match),
+    GET_SET_ENTRY(target_length),
+    GET_SET_ENTRY(strategy),
+    GET_SET_ENTRY(write_content_size),
+    GET_SET_ENTRY(write_checksum),
+    GET_SET_ENTRY(write_dict_id),
+    GET_SET_ENTRY(threads),
+    GET_SET_ENTRY(job_size),
+    GET_SET_ENTRY(overlap_log),
+    GET_SET_ENTRY(force_max_window),
+    GET_SET_ENTRY(enable_ldm),
+    GET_SET_ENTRY(ldm_hash_log),
+    GET_SET_ENTRY(ldm_min_match),
+    GET_SET_ENTRY(ldm_bucket_size_log),
+    GET_SET_ENTRY(ldm_hash_rate_log),
+    {NULL}};
+
+PyType_Slot ZstdCompressionParametersSlots[] = {
+    {Py_tp_dealloc, ZstdCompressionParameters_dealloc},
+    {Py_tp_methods, ZstdCompressionParameters_methods},
+    {Py_tp_getset, ZstdCompressionParameters_getset},
+    {Py_tp_init, ZstdCompressionParameters_init},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionParametersSpec = {
+    "zstd.ZstdCompressionParameters",
+    sizeof(ZstdCompressionParametersObject),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressionParametersSlots,
+};
+
+PyTypeObject *ZstdCompressionParametersType;
+
+void compressionparams_module_init(PyObject *mod) {
+    ZstdCompressionParametersType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionParametersSpec);
+    if (PyType_Ready(ZstdCompressionParametersType) < 0) {
+        return;
+    }
+
+    Py_INCREF(ZstdCompressionParametersType);
+    PyModule_AddObject(mod, "ZstdCompressionParameters",
+                       (PyObject *)ZstdCompressionParametersType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressionreader.c b/contrib/python/zstandard/py3/c-ext/compressionreader.c
new file mode 100644
index 00000000000..110ce9be866
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressionreader.c
@@ -0,0 +1,812 @@
+/**
+ * Copyright (c) 2017-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void compressionreader_dealloc(ZstdCompressionReader *self) {
+    Py_XDECREF(self->compressor);
+    Py_XDECREF(self->reader);
+
+    if (self->buffer.buf) {
+        PyBuffer_Release(&self->buffer);
+        memset(&self->buffer, 0, sizeof(self->buffer));
+    }
+
+    PyObject_Del(self);
+}
+
+static ZstdCompressionReader *
+compressionreader_enter(ZstdCompressionReader *self) {
+    if (self->entered) {
+        PyErr_SetString(PyExc_ValueError, "cannot __enter__ multiple times");
+        return NULL;
+    }
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    self->entered = 1;
+
+    Py_INCREF(self);
+    return self;
+}
+
+static PyObject *compressionreader_exit(ZstdCompressionReader *self,
+                                        PyObject *args) {
+    PyObject *exc_type;
+    PyObject *exc_value;
+    PyObject *exc_tb;
+    PyObject *result;
+
+    if (!PyArg_ParseTuple(args, "OOO:__exit__", &exc_type, &exc_value,
+                          &exc_tb)) {
+        return NULL;
+    }
+
+    self->entered = 0;
+
+    result = PyObject_CallMethod((PyObject *)self, "close", NULL);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    /* Release resources associated with source. */
+    Py_CLEAR(self->reader);
+    if (self->buffer.buf) {
+        PyBuffer_Release(&self->buffer);
+        memset(&self->buffer, 0, sizeof(self->buffer));
+    }
+
+    Py_CLEAR(self->compressor);
+
+    Py_RETURN_FALSE;
+}
+
+static PyObject *compressionreader_readable(ZstdCompressionReader *self) {
+    Py_RETURN_TRUE;
+}
+
+static PyObject *compressionreader_writable(ZstdCompressionReader *self) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *compressionreader_seekable(ZstdCompressionReader *self) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *compressionreader_readline(PyObject *self, PyObject *args) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *compressionreader_readlines(PyObject *self, PyObject *args) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *compressionreader_write(PyObject *self, PyObject *args) {
+    PyErr_SetString(PyExc_OSError, "stream is not writable");
+    return NULL;
+}
+
+static PyObject *compressionreader_writelines(PyObject *self, PyObject *args) {
+    PyErr_SetString(PyExc_OSError, "stream is not writable");
+    return NULL;
+}
+
+static PyObject *compressionreader_isatty(PyObject *self) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *compressionreader_flush(PyObject *self) {
+    Py_RETURN_NONE;
+}
+
+static PyObject *compressionreader_close(ZstdCompressionReader *self) {
+    if (self->closed) {
+        Py_RETURN_NONE;
+    }
+
+    self->closed = 1;
+
+    if (self->closefd && self->reader != NULL &&
+        PyObject_HasAttrString(self->reader, "close")) {
+        return PyObject_CallMethod(self->reader, "close", NULL);
+    }
+
+    Py_RETURN_NONE;
+}
+
+static PyObject *compressionreader_tell(ZstdCompressionReader *self) {
+    /* TODO should this raise OSError since stream isn't seekable? */
+    return PyLong_FromUnsignedLongLong(self->bytesCompressed);
+}
+
+int read_compressor_input(ZstdCompressionReader *self) {
+    if (self->finishedInput) {
+        return 0;
+    }
+
+    if (self->input.pos != self->input.size) {
+        return 0;
+    }
+
+    if (self->reader) {
+        Py_buffer buffer;
+
+        assert(self->readResult == NULL);
+
+        self->readResult =
+            PyObject_CallMethod(self->reader, "read", "k", self->readSize);
+
+        if (NULL == self->readResult) {
+            return -1;
+        }
+
+        memset(&buffer, 0, sizeof(buffer));
+
+        if (0 !=
+            PyObject_GetBuffer(self->readResult, &buffer, PyBUF_CONTIG_RO)) {
+            return -1;
+        }
+
+        /* EOF */
+        if (0 == buffer.len) {
+            self->finishedInput = 1;
+            Py_CLEAR(self->readResult);
+        }
+        else {
+            self->input.src = buffer.buf;
+            self->input.size = buffer.len;
+            self->input.pos = 0;
+        }
+
+        PyBuffer_Release(&buffer);
+    }
+    else {
+        assert(self->buffer.buf);
+
+        self->input.src = self->buffer.buf;
+        self->input.size = self->buffer.len;
+        self->input.pos = 0;
+    }
+
+    return 1;
+}
+
+int compress_input(ZstdCompressionReader *self, ZSTD_outBuffer *output) {
+    size_t oldPos;
+    size_t zresult;
+
+    /* If we have data left over, consume it. */
+    if (self->input.pos < self->input.size) {
+        oldPos = output->pos;
+
+        Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+            self->compressor->cctx, output, &self->input, ZSTD_e_continue);
+        Py_END_ALLOW_THREADS
+
+            self->bytesCompressed += output->pos - oldPos;
+
+        /* Input exhausted. Clear out state tracking. */
+        if (self->input.pos == self->input.size) {
+            memset(&self->input, 0, sizeof(self->input));
+            Py_CLEAR(self->readResult);
+
+            if (self->buffer.buf) {
+                self->finishedInput = 1;
+            }
+        }
+
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd compress error: %s",
+                         ZSTD_getErrorName(zresult));
+            return -1;
+        }
+    }
+
+    if (output->pos && output->pos == output->size) {
+        return 1;
+    }
+    else {
+        return 0;
+    }
+}
+
+static PyObject *compressionreader_read(ZstdCompressionReader *self,
+                                        PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"size", NULL};
+
+    Py_ssize_t size = -1;
+    PyObject *result = NULL;
+    char *resultBuffer;
+    Py_ssize_t resultSize;
+    size_t zresult;
+    size_t oldPos;
+    int readResult, compressResult;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n", kwlist, &size)) {
+        return NULL;
+    }
+
+    if (size < -1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot read negative amounts less than -1");
+        return NULL;
+    }
+
+    if (size == -1) {
+        return PyObject_CallMethod((PyObject *)self, "readall", NULL);
+    }
+
+    if (self->finishedOutput || size == 0) {
+        return PyBytes_FromStringAndSize("", 0);
+    }
+
+    result = PyBytes_FromStringAndSize(NULL, size);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+    self->output.dst = resultBuffer;
+    self->output.size = resultSize;
+    self->output.pos = 0;
+
+readinput:
+
+    compressResult = compress_input(self, &self->output);
+
+    if (-1 == compressResult) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+    else if (0 == compressResult) {
+        /* There is room in the output. We fall through to below, which will
+         * either get more input for us or will attempt to end the stream.
+         */
+    }
+    else if (1 == compressResult) {
+        memset(&self->output, 0, sizeof(self->output));
+        return result;
+    }
+    else {
+        assert(0);
+    }
+
+    readResult = read_compressor_input(self);
+
+    if (-1 == readResult) {
+        return NULL;
+    }
+    else if (0 == readResult) {
+    }
+    else if (1 == readResult) {
+    }
+    else {
+        assert(0);
+    }
+
+    if (self->input.size) {
+        goto readinput;
+    }
+
+    /* Else EOF */
+    oldPos = self->output.pos;
+
+    zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+                                   &self->input, ZSTD_e_end);
+
+    self->bytesCompressed += self->output.pos - oldPos;
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error ending compression stream: %s",
+                     ZSTD_getErrorName(zresult));
+        Py_XDECREF(result);
+        return NULL;
+    }
+
+    assert(self->output.pos);
+
+    if (0 == zresult) {
+        self->finishedOutput = 1;
+    }
+
+    if (safe_pybytes_resize(&result, self->output.pos)) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+
+    memset(&self->output, 0, sizeof(self->output));
+
+    return result;
+}
+
+static PyObject *compressionreader_read1(ZstdCompressionReader *self,
+                                         PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"size", NULL};
+
+    Py_ssize_t size = -1;
+    PyObject *result = NULL;
+    char *resultBuffer;
+    Py_ssize_t resultSize;
+    ZSTD_outBuffer output;
+    int compressResult;
+    size_t oldPos;
+    size_t zresult;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n:read1", kwlist, &size)) {
+        return NULL;
+    }
+
+    if (size < -1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot read negative amounts less than -1");
+        return NULL;
+    }
+
+    if (self->finishedOutput || size == 0) {
+        return PyBytes_FromStringAndSize("", 0);
+    }
+
+    if (size == -1) {
+        size = ZSTD_CStreamOutSize();
+    }
+
+    result = PyBytes_FromStringAndSize(NULL, size);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+    output.dst = resultBuffer;
+    output.size = resultSize;
+    output.pos = 0;
+
+    /* read1() is supposed to use at most 1 read() from the underlying stream.
+       However, we can't satisfy this requirement with compression because
+       not every input will generate output. We /could/ flush the compressor,
+       but this may not be desirable. We allow multiple read() from the
+       underlying stream. But unlike read(), we return as soon as output data
+       is available.
+    */
+
+    compressResult = compress_input(self, &output);
+
+    if (-1 == compressResult) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+    else if (0 == compressResult || 1 == compressResult) {
+    }
+    else {
+        assert(0);
+    }
+
+    if (output.pos) {
+        goto finally;
+    }
+
+    while (!self->finishedInput) {
+        int readResult = read_compressor_input(self);
+
+        if (-1 == readResult) {
+            Py_XDECREF(result);
+            return NULL;
+        }
+        else if (0 == readResult || 1 == readResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        compressResult = compress_input(self, &output);
+
+        if (-1 == compressResult) {
+            Py_XDECREF(result);
+            return NULL;
+        }
+        else if (0 == compressResult || 1 == compressResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        if (output.pos) {
+            goto finally;
+        }
+    }
+
+    /* EOF */
+    oldPos = output.pos;
+
+    zresult = ZSTD_compressStream2(self->compressor->cctx, &output,
+                                   &self->input, ZSTD_e_end);
+
+    self->bytesCompressed += output.pos - oldPos;
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error ending compression stream: %s",
+                     ZSTD_getErrorName(zresult));
+        Py_XDECREF(result);
+        return NULL;
+    }
+
+    if (zresult == 0) {
+        self->finishedOutput = 1;
+    }
+
+finally:
+    if (result) {
+        if (safe_pybytes_resize(&result, output.pos)) {
+            Py_XDECREF(result);
+            return NULL;
+        }
+    }
+
+    return result;
+}
+
+static PyObject *compressionreader_readall(PyObject *self) {
+    PyObject *chunks = NULL;
+    PyObject *empty = NULL;
+    PyObject *result = NULL;
+
+    /* Our strategy is to collect chunks into a list then join all the
+     * chunks at the end. We could potentially use e.g. an io.BytesIO. But
+     * this feels simple enough to implement and avoids potentially expensive
+     * reallocations of large buffers.
+     */
+    chunks = PyList_New(0);
+    if (NULL == chunks) {
+        return NULL;
+    }
+
+    while (1) {
+        PyObject *chunk = PyObject_CallMethod(self, "read", "i", 1048576);
+        if (NULL == chunk) {
+            Py_DECREF(chunks);
+            return NULL;
+        }
+
+        if (!PyBytes_Size(chunk)) {
+            Py_DECREF(chunk);
+            break;
+        }
+
+        if (PyList_Append(chunks, chunk)) {
+            Py_DECREF(chunk);
+            Py_DECREF(chunks);
+            return NULL;
+        }
+
+        Py_DECREF(chunk);
+    }
+
+    empty = PyBytes_FromStringAndSize("", 0);
+    if (NULL == empty) {
+        Py_DECREF(chunks);
+        return NULL;
+    }
+
+    result = PyObject_CallMethod(empty, "join", "O", chunks);
+
+    Py_DECREF(empty);
+    Py_DECREF(chunks);
+
+    return result;
+}
+
+static PyObject *compressionreader_readinto(ZstdCompressionReader *self,
+                                            PyObject *args) {
+    Py_buffer dest;
+    ZSTD_outBuffer output;
+    int readResult, compressResult;
+    PyObject *result = NULL;
+    size_t zresult;
+    size_t oldPos;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (self->finishedOutput) {
+        return PyLong_FromLong(0);
+    }
+
+    if (!PyArg_ParseTuple(args, "w*:readinto", &dest)) {
+        return NULL;
+    }
+
+    output.dst = dest.buf;
+    output.size = dest.len;
+    output.pos = 0;
+
+    compressResult = compress_input(self, &output);
+
+    if (-1 == compressResult) {
+        goto finally;
+    }
+    else if (0 == compressResult) {
+    }
+    else if (1 == compressResult) {
+        result = PyLong_FromSize_t(output.pos);
+        goto finally;
+    }
+    else {
+        assert(0);
+    }
+
+    while (!self->finishedInput) {
+        readResult = read_compressor_input(self);
+
+        if (-1 == readResult) {
+            goto finally;
+        }
+        else if (0 == readResult || 1 == readResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        compressResult = compress_input(self, &output);
+
+        if (-1 == compressResult) {
+            goto finally;
+        }
+        else if (0 == compressResult) {
+        }
+        else if (1 == compressResult) {
+            result = PyLong_FromSize_t(output.pos);
+            goto finally;
+        }
+        else {
+            assert(0);
+        }
+    }
+
+    /* EOF */
+    oldPos = output.pos;
+
+    zresult = ZSTD_compressStream2(self->compressor->cctx, &output,
+                                   &self->input, ZSTD_e_end);
+
+    self->bytesCompressed += self->output.pos - oldPos;
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error ending compression stream: %s",
+                     ZSTD_getErrorName(zresult));
+        goto finally;
+    }
+
+    assert(output.pos);
+
+    if (0 == zresult) {
+        self->finishedOutput = 1;
+    }
+
+    result = PyLong_FromSize_t(output.pos);
+
+finally:
+    PyBuffer_Release(&dest);
+
+    return result;
+}
+
+static PyObject *compressionreader_readinto1(ZstdCompressionReader *self,
+                                             PyObject *args) {
+    Py_buffer dest;
+    PyObject *result = NULL;
+    ZSTD_outBuffer output;
+    int compressResult;
+    size_t oldPos;
+    size_t zresult;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (self->finishedOutput) {
+        return PyLong_FromLong(0);
+    }
+
+    if (!PyArg_ParseTuple(args, "w*:readinto1", &dest)) {
+        return NULL;
+    }
+
+    output.dst = dest.buf;
+    output.size = dest.len;
+    output.pos = 0;
+
+    compressResult = compress_input(self, &output);
+
+    if (-1 == compressResult) {
+        goto finally;
+    }
+    else if (0 == compressResult || 1 == compressResult) {
+    }
+    else {
+        assert(0);
+    }
+
+    if (output.pos) {
+        result = PyLong_FromSize_t(output.pos);
+        goto finally;
+    }
+
+    while (!self->finishedInput) {
+        int readResult = read_compressor_input(self);
+
+        if (-1 == readResult) {
+            goto finally;
+        }
+        else if (0 == readResult || 1 == readResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        compressResult = compress_input(self, &output);
+
+        if (-1 == compressResult) {
+            goto finally;
+        }
+        else if (0 == compressResult) {
+        }
+        else if (1 == compressResult) {
+            result = PyLong_FromSize_t(output.pos);
+            goto finally;
+        }
+        else {
+            assert(0);
+        }
+
+        /* If we produced output and we're not done with input, emit
+         * that output now, as we've hit restrictions of read1().
+         */
+        if (output.pos && !self->finishedInput) {
+            result = PyLong_FromSize_t(output.pos);
+            goto finally;
+        }
+
+        /* Otherwise we either have no output or we've exhausted the
+         * input. Either we try to get more input or we fall through
+         * to EOF below */
+    }
+
+    /* EOF */
+    oldPos = output.pos;
+
+    zresult = ZSTD_compressStream2(self->compressor->cctx, &output,
+                                   &self->input, ZSTD_e_end);
+
+    self->bytesCompressed += self->output.pos - oldPos;
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error ending compression stream: %s",
+                     ZSTD_getErrorName(zresult));
+        goto finally;
+    }
+
+    assert(output.pos);
+
+    if (0 == zresult) {
+        self->finishedOutput = 1;
+    }
+
+    result = PyLong_FromSize_t(output.pos);
+
+finally:
+    PyBuffer_Release(&dest);
+
+    return result;
+}
+
+static PyObject *compressionreader_iter(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *compressionreader_iternext(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyMethodDef compressionreader_methods[] = {
+    {"__enter__", (PyCFunction)compressionreader_enter, METH_NOARGS,
+     PyDoc_STR("Enter a compression context")},
+    {"__exit__", (PyCFunction)compressionreader_exit, METH_VARARGS,
+     PyDoc_STR("Exit a compression context")},
+    {"close", (PyCFunction)compressionreader_close, METH_NOARGS,
+     PyDoc_STR("Close the stream so it cannot perform any more operations")},
+    {"flush", (PyCFunction)compressionreader_flush, METH_NOARGS,
+     PyDoc_STR("no-ops")},
+    {"isatty", (PyCFunction)compressionreader_isatty, METH_NOARGS,
+     PyDoc_STR("Returns False")},
+    {"readable", (PyCFunction)compressionreader_readable, METH_NOARGS,
+     PyDoc_STR("Returns True")},
+    {"read", (PyCFunction)compressionreader_read, METH_VARARGS | METH_KEYWORDS,
+     PyDoc_STR("read compressed data")},
+    {"read1", (PyCFunction)compressionreader_read1,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readall", (PyCFunction)compressionreader_readall, METH_NOARGS,
+     PyDoc_STR("Not implemented")},
+    {"readinto", (PyCFunction)compressionreader_readinto, METH_VARARGS, NULL},
+    {"readinto1", (PyCFunction)compressionreader_readinto1, METH_VARARGS, NULL},
+    {"readline", (PyCFunction)compressionreader_readline, METH_VARARGS,
+     PyDoc_STR("Not implemented")},
+    {"readlines", (PyCFunction)compressionreader_readlines, METH_VARARGS,
+     PyDoc_STR("Not implemented")},
+    {"seekable", (PyCFunction)compressionreader_seekable, METH_NOARGS,
+     PyDoc_STR("Returns False")},
+    {"tell", (PyCFunction)compressionreader_tell, METH_NOARGS,
+     PyDoc_STR("Returns current number of bytes compressed")},
+    {"writable", (PyCFunction)compressionreader_writable, METH_NOARGS,
+     PyDoc_STR("Returns False")},
+    {"write", compressionreader_write, METH_VARARGS,
+     PyDoc_STR("Raises OSError")},
+    {"writelines", compressionreader_writelines, METH_VARARGS,
+     PyDoc_STR("Not implemented")},
+    {NULL, NULL}};
+
+static PyMemberDef compressionreader_members[] = {
+    {"closed", T_BOOL, offsetof(ZstdCompressionReader, closed), READONLY,
+     "whether stream is closed"},
+    {NULL}};
+
+PyType_Slot ZstdCompressionReaderSlots[] = {
+    {Py_tp_dealloc, compressionreader_dealloc},
+    {Py_tp_iter, compressionreader_iter},
+    {Py_tp_iternext, compressionreader_iternext},
+    {Py_tp_methods, compressionreader_methods},
+    {Py_tp_members, compressionreader_members},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionReaderSpec = {
+    "zstd.ZstdCompressionReader",
+    sizeof(ZstdCompressionReader),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    ZstdCompressionReaderSlots,
+};
+
+PyTypeObject *ZstdCompressionReaderType;
+
+void compressionreader_module_init(PyObject *mod) {
+    /* TODO make reader a sub-class of io.RawIOBase */
+
+    ZstdCompressionReaderType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionReaderSpec);
+    if (PyType_Ready(ZstdCompressionReaderType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdCompressionReaderType);
+    PyModule_AddObject(mod, "ZstdCompressionReader",
+                       (PyObject *)ZstdCompressionReaderType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressionwriter.c b/contrib/python/zstandard/py3/c-ext/compressionwriter.c
new file mode 100644
index 00000000000..1def1724b9b
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressionwriter.c
@@ -0,0 +1,352 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void ZstdCompressionWriter_dealloc(ZstdCompressionWriter *self) {
+    Py_XDECREF(self->compressor);
+    Py_XDECREF(self->writer);
+
+    PyMem_Free(self->output.dst);
+    self->output.dst = NULL;
+
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdCompressionWriter_enter(ZstdCompressionWriter *self) {
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (self->entered) {
+        PyErr_SetString(ZstdError, "cannot __enter__ multiple times");
+        return NULL;
+    }
+
+    self->entered = 1;
+
+    Py_INCREF(self);
+    return (PyObject *)self;
+}
+
+static PyObject *ZstdCompressionWriter_exit(ZstdCompressionWriter *self,
+                                            PyObject *args) {
+    PyObject *exc_type;
+    PyObject *exc_value;
+    PyObject *exc_tb;
+
+    if (!PyArg_ParseTuple(args, "OOO:__exit__", &exc_type, &exc_value,
+                          &exc_tb)) {
+        return NULL;
+    }
+
+    self->entered = 0;
+
+    PyObject *result = PyObject_CallMethod((PyObject *)self, "close", NULL);
+
+    if (NULL == result) {
+        return NULL;
+    }
+
+    Py_RETURN_FALSE;
+}
+
+static PyObject *
+ZstdCompressionWriter_memory_size(ZstdCompressionWriter *self) {
+    return PyLong_FromSize_t(ZSTD_sizeof_CCtx(self->compressor->cctx));
+}
+
+static PyObject *ZstdCompressionWriter_write(ZstdCompressionWriter *self,
+                                             PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    PyObject *result = NULL;
+    Py_buffer source;
+    size_t zresult;
+    ZSTD_inBuffer input;
+    PyObject *res;
+    Py_ssize_t totalWrite = 0;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:write", kwlist,
+                                     &source)) {
+        return NULL;
+    }
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    self->output.pos = 0;
+
+    input.src = source.buf;
+    input.size = source.len;
+    input.pos = 0;
+
+    while (input.pos < (size_t)source.len) {
+        Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+            self->compressor->cctx, &self->output, &input, ZSTD_e_continue);
+        Py_END_ALLOW_THREADS
+
+            if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd compress error: %s",
+                         ZSTD_getErrorName(zresult));
+            goto finally;
+        }
+
+        /* Copy data from output buffer to writer. */
+        if (self->output.pos) {
+            res = PyObject_CallMethod(self->writer, "write", "y#",
+                                      self->output.dst, self->output.pos);
+            if (NULL == res) {
+                goto finally;
+            }
+            Py_XDECREF(res);
+            totalWrite += self->output.pos;
+            self->bytesCompressed += self->output.pos;
+        }
+        self->output.pos = 0;
+    }
+
+    if (self->writeReturnRead) {
+        result = PyLong_FromSize_t(input.pos);
+    }
+    else {
+        result = PyLong_FromSsize_t(totalWrite);
+    }
+
+finally:
+    PyBuffer_Release(&source);
+    return result;
+}
+
+static PyObject *ZstdCompressionWriter_flush(ZstdCompressionWriter *self,
+                                             PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"flush_mode", NULL};
+
+    size_t zresult;
+    ZSTD_inBuffer input;
+    PyObject *res;
+    Py_ssize_t totalWrite = 0;
+    unsigned flush_mode = 0;
+    ZSTD_EndDirective flush;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|I:flush", kwlist,
+                                     &flush_mode)) {
+        return NULL;
+    }
+
+    switch (flush_mode) {
+    case 0:
+        flush = ZSTD_e_flush;
+        break;
+    case 1:
+        flush = ZSTD_e_end;
+        break;
+    default:
+        PyErr_Format(PyExc_ValueError, "unknown flush_mode: %d", flush_mode);
+        return NULL;
+    }
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    self->output.pos = 0;
+
+    input.src = NULL;
+    input.size = 0;
+    input.pos = 0;
+
+    while (1) {
+        Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+            self->compressor->cctx, &self->output, &input, flush);
+        Py_END_ALLOW_THREADS
+
+            if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd compress error: %s",
+                         ZSTD_getErrorName(zresult));
+            return NULL;
+        }
+
+        /* Copy data from output buffer to writer. */
+        if (self->output.pos) {
+            res = PyObject_CallMethod(self->writer, "write", "y#",
+                                      self->output.dst, self->output.pos);
+            if (NULL == res) {
+                return NULL;
+            }
+            Py_XDECREF(res);
+            totalWrite += self->output.pos;
+            self->bytesCompressed += self->output.pos;
+        }
+
+        self->output.pos = 0;
+
+        if (!zresult) {
+            break;
+        }
+    }
+
+    if (!self->closing && PyObject_HasAttrString(self->writer, "flush")) {
+        res = PyObject_CallMethod(self->writer, "flush", NULL);
+        if (NULL == res) {
+            return NULL;
+        }
+        Py_XDECREF(res);
+    }
+
+    return PyLong_FromSsize_t(totalWrite);
+}
+
+static PyObject *ZstdCompressionWriter_close(ZstdCompressionWriter *self) {
+    PyObject *result;
+
+    if (self->closed) {
+        Py_RETURN_NONE;
+    }
+
+    self->closing = 1;
+    result = PyObject_CallMethod((PyObject *)self, "flush", "I", 1);
+    self->closing = 0;
+    self->closed = 1;
+
+    if (NULL == result) {
+        return NULL;
+    }
+
+    /* Call close on underlying stream as well. */
+    if (self->closefd && PyObject_HasAttrString(self->writer, "close")) {
+        return PyObject_CallMethod(self->writer, "close", NULL);
+    }
+
+    Py_RETURN_NONE;
+}
+
+static PyObject *ZstdCompressionWriter_fileno(ZstdCompressionWriter *self) {
+    if (PyObject_HasAttrString(self->writer, "fileno")) {
+        return PyObject_CallMethod(self->writer, "fileno", NULL);
+    }
+    else {
+        PyErr_SetString(PyExc_OSError,
+                        "fileno not available on underlying writer");
+        return NULL;
+    }
+}
+
+static PyObject *ZstdCompressionWriter_tell(ZstdCompressionWriter *self) {
+    return PyLong_FromUnsignedLongLong(self->bytesCompressed);
+}
+
+static PyObject *ZstdCompressionWriter_writelines(PyObject *self,
+                                                  PyObject *args) {
+    PyErr_SetNone(PyExc_NotImplementedError);
+    return NULL;
+}
+
+static PyObject *ZstdCompressionWriter_iter(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *ZstdCompressionWriter_iternext(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *ZstdCompressionWriter_false(PyObject *self, PyObject *args) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *ZstdCompressionWriter_true(PyObject *self, PyObject *args) {
+    Py_RETURN_TRUE;
+}
+
+static PyObject *ZstdCompressionWriter_unsupported(PyObject *self,
+                                                   PyObject *args,
+                                                   PyObject *kwargs) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyMethodDef ZstdCompressionWriter_methods[] = {
+    {"__enter__", (PyCFunction)ZstdCompressionWriter_enter, METH_NOARGS,
+     PyDoc_STR("Enter a compression context.")},
+    {"__exit__", (PyCFunction)ZstdCompressionWriter_exit, METH_VARARGS,
+     PyDoc_STR("Exit a compression context.")},
+    {"close", (PyCFunction)ZstdCompressionWriter_close, METH_NOARGS, NULL},
+    {"fileno", (PyCFunction)ZstdCompressionWriter_fileno, METH_NOARGS, NULL},
+    {"isatty", (PyCFunction)ZstdCompressionWriter_false, METH_NOARGS, NULL},
+    {"readable", (PyCFunction)ZstdCompressionWriter_false, METH_NOARGS, NULL},
+    {"readline", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readlines", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"seek", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"seekable", ZstdCompressionWriter_false, METH_NOARGS, NULL},
+    {"truncate", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"writable", ZstdCompressionWriter_true, METH_NOARGS, NULL},
+    {"writelines", ZstdCompressionWriter_writelines, METH_VARARGS, NULL},
+    {"read", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readall", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readinto", (PyCFunction)ZstdCompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"memory_size", (PyCFunction)ZstdCompressionWriter_memory_size, METH_NOARGS,
+     PyDoc_STR("Obtain the memory size of the underlying compressor")},
+    {"write", (PyCFunction)ZstdCompressionWriter_write,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("Compress data")},
+    {"flush", (PyCFunction)ZstdCompressionWriter_flush,
+     METH_VARARGS | METH_KEYWORDS,
+     PyDoc_STR("Flush data and finish a zstd frame")},
+    {"tell", (PyCFunction)ZstdCompressionWriter_tell, METH_NOARGS,
+     PyDoc_STR("Returns current number of bytes compressed")},
+    {NULL, NULL}};
+
+static PyMemberDef ZstdCompressionWriter_members[] = {
+    {"closed", T_BOOL, offsetof(ZstdCompressionWriter, closed), READONLY, NULL},
+    {NULL}};
+
+PyType_Slot ZstdCompressionWriterSlots[] = {
+    {Py_tp_dealloc, ZstdCompressionWriter_dealloc},
+    {Py_tp_iter, ZstdCompressionWriter_iter},
+    {Py_tp_iternext, ZstdCompressionWriter_iternext},
+    {Py_tp_methods, ZstdCompressionWriter_methods},
+    {Py_tp_members, ZstdCompressionWriter_members},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionWriterSpec = {
+    "zstd.ZstdCompressionWriter",
+    sizeof(ZstdCompressionWriter),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressionWriterSlots,
+};
+
+PyTypeObject *ZstdCompressionWriterType;
+
+void compressionwriter_module_init(PyObject *mod) {
+    ZstdCompressionWriterType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionWriterSpec);
+    if (PyType_Ready(ZstdCompressionWriterType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdCompressionWriterType);
+    PyModule_AddObject(mod, "ZstdCompressionWriter",
+                       (PyObject *)ZstdCompressionWriterType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressobj.c b/contrib/python/zstandard/py3/c-ext/compressobj.c
new file mode 100644
index 00000000000..5e224475b45
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressobj.c
@@ -0,0 +1,220 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void ZstdCompressionObj_dealloc(ZstdCompressionObj *self) {
+    PyMem_Free(self->output.dst);
+    self->output.dst = NULL;
+
+    Py_XDECREF(self->compressor);
+
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdCompressionObj_compress(ZstdCompressionObj *self,
+                                             PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    Py_buffer source;
+    ZSTD_inBuffer input;
+    size_t zresult;
+    PyObject *result = NULL;
+    Py_ssize_t resultSize = 0;
+
+    if (self->finished) {
+        PyErr_SetString(ZstdError,
+                        "cannot call compress() after compressor finished");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:compress", kwlist,
+                                     &source)) {
+        return NULL;
+    }
+
+    input.src = source.buf;
+    input.size = source.len;
+    input.pos = 0;
+
+    while (input.pos < (size_t)source.len) {
+        Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+            self->compressor->cctx, &self->output, &input, ZSTD_e_continue);
+        Py_END_ALLOW_THREADS
+
+            if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd compress error: %s",
+                         ZSTD_getErrorName(zresult));
+            Py_CLEAR(result);
+            goto finally;
+        }
+
+        if (self->output.pos) {
+            if (result) {
+                resultSize = PyBytes_GET_SIZE(result);
+
+                if (safe_pybytes_resize(&result,
+                                        resultSize + self->output.pos)) {
+                    Py_CLEAR(result);
+                    goto finally;
+                }
+
+                memcpy(PyBytes_AS_STRING(result) + resultSize, self->output.dst,
+                       self->output.pos);
+            }
+            else {
+                result = PyBytes_FromStringAndSize(self->output.dst,
+                                                   self->output.pos);
+                if (!result) {
+                    goto finally;
+                }
+            }
+
+            self->output.pos = 0;
+        }
+    }
+
+    if (NULL == result) {
+        result = PyBytes_FromString("");
+    }
+
+finally:
+    PyBuffer_Release(&source);
+
+    return result;
+}
+
+static PyObject *ZstdCompressionObj_flush(ZstdCompressionObj *self,
+                                          PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"flush_mode", NULL};
+
+    int flushMode = compressorobj_flush_finish;
+    size_t zresult;
+    PyObject *result = NULL;
+    Py_ssize_t resultSize = 0;
+    ZSTD_inBuffer input;
+    ZSTD_EndDirective zFlushMode;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|i:flush", kwlist,
+                                     &flushMode)) {
+        return NULL;
+    }
+
+    if (flushMode != compressorobj_flush_finish &&
+        flushMode != compressorobj_flush_block) {
+        PyErr_SetString(PyExc_ValueError, "flush mode not recognized");
+        return NULL;
+    }
+
+    if (self->finished) {
+        PyErr_SetString(ZstdError, "compressor object already finished");
+        return NULL;
+    }
+
+    switch (flushMode) {
+    case compressorobj_flush_block:
+        zFlushMode = ZSTD_e_flush;
+        break;
+
+    case compressorobj_flush_finish:
+        zFlushMode = ZSTD_e_end;
+        self->finished = 1;
+        break;
+
+    default:
+        PyErr_SetString(ZstdError, "unhandled flush mode");
+        return NULL;
+    }
+
+    assert(self->output.pos == 0);
+
+    input.src = NULL;
+    input.size = 0;
+    input.pos = 0;
+
+    while (1) {
+        Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+            self->compressor->cctx, &self->output, &input, zFlushMode);
+        Py_END_ALLOW_THREADS
+
+            if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "error ending compression stream: %s",
+                         ZSTD_getErrorName(zresult));
+            return NULL;
+        }
+
+        if (self->output.pos) {
+            if (result) {
+                resultSize = PyBytes_GET_SIZE(result);
+
+                if (safe_pybytes_resize(&result,
+                                        resultSize + self->output.pos)) {
+                    Py_XDECREF(result);
+                    return NULL;
+                }
+
+                memcpy(PyBytes_AS_STRING(result) + resultSize, self->output.dst,
+                       self->output.pos);
+            }
+            else {
+                result = PyBytes_FromStringAndSize(self->output.dst,
+                                                   self->output.pos);
+                if (!result) {
+                    return NULL;
+                }
+            }
+
+            self->output.pos = 0;
+        }
+
+        if (!zresult) {
+            break;
+        }
+    }
+
+    if (result) {
+        return result;
+    }
+    else {
+        return PyBytes_FromString("");
+    }
+}
+
+static PyMethodDef ZstdCompressionObj_methods[] = {
+    {"compress", (PyCFunction)ZstdCompressionObj_compress,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("compress data")},
+    {"flush", (PyCFunction)ZstdCompressionObj_flush,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("finish compression operation")},
+    {NULL, NULL}};
+
+PyType_Slot ZstdCompressionObjSlots[] = {
+    {Py_tp_dealloc, ZstdCompressionObj_dealloc},
+    {Py_tp_methods, ZstdCompressionObj_methods},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressionObjSpec = {
+    "zstd.ZstdCompressionObj",
+    sizeof(ZstdCompressionObj),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressionObjSlots,
+};
+
+PyTypeObject *ZstdCompressionObjType;
+
+void compressobj_module_init(PyObject *module) {
+    ZstdCompressionObjType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressionObjSpec);
+    if (PyType_Ready(ZstdCompressionObjType) < 0) {
+        return;
+    }
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressor.c b/contrib/python/zstandard/py3/c-ext/compressor.c
new file mode 100644
index 00000000000..bbef9fd79eb
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressor.c
@@ -0,0 +1,1557 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+int setup_cctx(ZstdCompressor *compressor) {
+    size_t zresult;
+
+    assert(compressor);
+    assert(compressor->cctx);
+    assert(compressor->params);
+
+    zresult = ZSTD_CCtx_setParametersUsingCCtxParams(compressor->cctx,
+                                                     compressor->params);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "could not set compression parameters: %s",
+                     ZSTD_getErrorName(zresult));
+        return 1;
+    }
+
+    if (compressor->dict) {
+        if (compressor->dict->cdict) {
+            zresult =
+                ZSTD_CCtx_refCDict(compressor->cctx, compressor->dict->cdict);
+        }
+        else {
+            zresult = ZSTD_CCtx_loadDictionary_advanced(
+                compressor->cctx, compressor->dict->dictData,
+                compressor->dict->dictSize, ZSTD_dlm_byRef,
+                compressor->dict->dictType);
+        }
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "could not load compression dictionary: %s",
+                         ZSTD_getErrorName(zresult));
+            return 1;
+        }
+    }
+
+    return 0;
+}
+
+static PyObject *frame_progression(ZSTD_CCtx *cctx) {
+    PyObject *result = NULL;
+    PyObject *value;
+    ZSTD_frameProgression progression;
+
+    result = PyTuple_New(3);
+    if (!result) {
+        return NULL;
+    }
+
+    progression = ZSTD_getFrameProgression(cctx);
+
+    value = PyLong_FromUnsignedLongLong(progression.ingested);
+    if (!value) {
+        Py_DECREF(result);
+        return NULL;
+    }
+
+    PyTuple_SET_ITEM(result, 0, value);
+
+    value = PyLong_FromUnsignedLongLong(progression.consumed);
+    if (!value) {
+        Py_DECREF(result);
+        return NULL;
+    }
+
+    PyTuple_SET_ITEM(result, 1, value);
+
+    value = PyLong_FromUnsignedLongLong(progression.produced);
+    if (!value) {
+        Py_DECREF(result);
+        return NULL;
+    }
+
+    PyTuple_SET_ITEM(result, 2, value);
+
+    return result;
+}
+
+static int ZstdCompressor_init(ZstdCompressor *self, PyObject *args,
+                               PyObject *kwargs) {
+    static char *kwlist[] = {"level",
+                             "dict_data",
+                             "compression_params",
+                             "write_checksum",
+                             "write_content_size",
+                             "write_dict_id",
+                             "threads",
+                             NULL};
+
+    int level = 3;
+    PyObject *dict = NULL;
+    PyObject *params = NULL;
+    PyObject *writeChecksum = NULL;
+    PyObject *writeContentSize = NULL;
+    PyObject *writeDictID = NULL;
+    int threads = 0;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|iOOOOOi:ZstdCompressor",
+                                     kwlist, &level, &dict, &params,
+                                     &writeChecksum, &writeContentSize,
+                                     &writeDictID, &threads)) {
+        return -1;
+    }
+
+    if (level > ZSTD_maxCLevel()) {
+        PyErr_Format(PyExc_ValueError, "level must be less than %d",
+                     ZSTD_maxCLevel() + 1);
+        return -1;
+    }
+
+    if (threads < 0) {
+        threads = cpu_count();
+    }
+
+    if (dict) {
+        if (dict == Py_None) {
+            dict = NULL;
+        }
+        else if (!PyObject_IsInstance(dict,
+                                      (PyObject *)ZstdCompressionDictType)) {
+            PyErr_Format(PyExc_TypeError,
+                         "dict_data must be zstd.ZstdCompressionDict");
+            return -1;
+        }
+    }
+
+    if (params) {
+        if (params == Py_None) {
+            params = NULL;
+        }
+        else if (!PyObject_IsInstance(
+                     params, (PyObject *)ZstdCompressionParametersType)) {
+            PyErr_Format(
+                PyExc_TypeError,
+                "compression_params must be zstd.ZstdCompressionParameters");
+            return -1;
+        }
+    }
+
+    if (writeChecksum == Py_None) {
+        writeChecksum = NULL;
+    }
+    if (writeContentSize == Py_None) {
+        writeContentSize = NULL;
+    }
+    if (writeDictID == Py_None) {
+        writeDictID = NULL;
+    }
+
+    /* We create a ZSTD_CCtx for reuse among multiple operations to reduce the
+       overhead of each compression operation. */
+    self->cctx = ZSTD_createCCtx();
+    if (!self->cctx) {
+        PyErr_NoMemory();
+        return -1;
+    }
+
+    /* TODO stuff the original parameters away somewhere so we can reset later.
+       This will allow us to do things like automatically adjust cparams based
+       on input size (assuming zstd isn't doing that internally). */
+
+    self->params = ZSTD_createCCtxParams();
+    if (!self->params) {
+        PyErr_NoMemory();
+        return -1;
+    }
+
+    if (params && writeChecksum) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot define compression_params and write_checksum");
+        return -1;
+    }
+
+    if (params && writeContentSize) {
+        PyErr_SetString(
+            PyExc_ValueError,
+            "cannot define compression_params and write_content_size");
+        return -1;
+    }
+
+    if (params && writeDictID) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot define compression_params and write_dict_id");
+        return -1;
+    }
+
+    if (params && threads) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot define compression_params and threads");
+        return -1;
+    }
+
+    if (params) {
+        if (set_parameters(self->params,
+                           (ZstdCompressionParametersObject *)params)) {
+            return -1;
+        }
+    }
+    else {
+        if (set_parameter(self->params, ZSTD_c_compressionLevel, level)) {
+            return -1;
+        }
+
+        if (set_parameter(self->params, ZSTD_c_contentSizeFlag,
+                          writeContentSize ? PyObject_IsTrue(writeContentSize)
+                                           : 1)) {
+            return -1;
+        }
+
+        if (set_parameter(self->params, ZSTD_c_checksumFlag,
+                          writeChecksum ? PyObject_IsTrue(writeChecksum) : 0)) {
+            return -1;
+        }
+
+        if (set_parameter(self->params, ZSTD_c_dictIDFlag,
+                          writeDictID ? PyObject_IsTrue(writeDictID) : 1)) {
+            return -1;
+        }
+
+        if (threads) {
+            if (set_parameter(self->params, ZSTD_c_nbWorkers, threads)) {
+                return -1;
+            }
+        }
+    }
+
+    if (dict) {
+        self->dict = (ZstdCompressionDict *)dict;
+        Py_INCREF(dict);
+    }
+
+    if (setup_cctx(self)) {
+        return -1;
+    }
+
+    return 0;
+}
+
+static void ZstdCompressor_dealloc(ZstdCompressor *self) {
+    if (self->cctx) {
+        ZSTD_freeCCtx(self->cctx);
+        self->cctx = NULL;
+    }
+
+    if (self->params) {
+        ZSTD_freeCCtxParams(self->params);
+        self->params = NULL;
+    }
+
+    Py_XDECREF(self->dict);
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdCompressor_memory_size(ZstdCompressor *self) {
+    if (self->cctx) {
+        return PyLong_FromSize_t(ZSTD_sizeof_CCtx(self->cctx));
+    }
+    else {
+        PyErr_SetString(
+            ZstdError, "no compressor context found; this should never happen");
+        return NULL;
+    }
+}
+
+static PyObject *ZstdCompressor_frame_progression(ZstdCompressor *self) {
+    return frame_progression(self->cctx);
+}
+
+static PyObject *ZstdCompressor_copy_stream(ZstdCompressor *self,
+                                            PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"ifh",       "ofh",        "size",
+                             "read_size", "write_size", NULL};
+
+    PyObject *source;
+    PyObject *dest;
+    unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    size_t inSize = ZSTD_CStreamInSize();
+    size_t outSize = ZSTD_CStreamOutSize();
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    Py_ssize_t totalRead = 0;
+    Py_ssize_t totalWrite = 0;
+    char *readBuffer;
+    Py_ssize_t readSize;
+    PyObject *readResult = NULL;
+    PyObject *res = NULL;
+    size_t zresult;
+    PyObject *writeResult;
+    PyObject *totalReadPy;
+    PyObject *totalWritePy;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "OO|Kkk:copy_stream", kwlist,
+                                     &source, &dest, &sourceSize, &inSize,
+                                     &outSize)) {
+        return NULL;
+    }
+
+    if (!PyObject_HasAttrString(source, "read")) {
+        PyErr_SetString(PyExc_ValueError,
+                        "first argument must have a read() method");
+        return NULL;
+    }
+
+    if (!PyObject_HasAttrString(dest, "write")) {
+        PyErr_SetString(PyExc_ValueError,
+                        "second argument must have a write() method");
+        return NULL;
+    }
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source size: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    /* Prevent free on uninitialized memory in finally. */
+    output.dst = PyMem_Malloc(outSize);
+    if (!output.dst) {
+        PyErr_NoMemory();
+        res = NULL;
+        goto finally;
+    }
+    output.size = outSize;
+    output.pos = 0;
+
+    input.src = NULL;
+    input.size = 0;
+    input.pos = 0;
+
+    while (1) {
+        /* Try to read from source stream. */
+        readResult = PyObject_CallMethod(source, "read", "n", inSize);
+        if (!readResult) {
+            goto finally;
+        }
+
+        PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+
+        /* If no data was read, we're at EOF. */
+        if (0 == readSize) {
+            break;
+        }
+
+        totalRead += readSize;
+
+        /* Send data to compressor */
+        input.src = readBuffer;
+        input.size = readSize;
+        input.pos = 0;
+
+        while (input.pos < input.size) {
+            Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+                self->cctx, &output, &input, ZSTD_e_continue);
+            Py_END_ALLOW_THREADS
+
+                if (ZSTD_isError(zresult)) {
+                res = NULL;
+                PyErr_Format(ZstdError, "zstd compress error: %s",
+                             ZSTD_getErrorName(zresult));
+                goto finally;
+            }
+
+            if (output.pos) {
+                writeResult = PyObject_CallMethod(dest, "write", "y#",
+                                                  output.dst, output.pos);
+                if (NULL == writeResult) {
+                    res = NULL;
+                    goto finally;
+                }
+                Py_XDECREF(writeResult);
+                totalWrite += output.pos;
+                output.pos = 0;
+            }
+        }
+
+        Py_CLEAR(readResult);
+    }
+
+    /* We've finished reading. Now flush the compressor stream. */
+    assert(input.pos == input.size);
+
+    while (1) {
+        Py_BEGIN_ALLOW_THREADS zresult =
+            ZSTD_compressStream2(self->cctx, &output, &input, ZSTD_e_end);
+        Py_END_ALLOW_THREADS
+
+            if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "error ending compression stream: %s",
+                         ZSTD_getErrorName(zresult));
+            res = NULL;
+            goto finally;
+        }
+
+        if (output.pos) {
+            writeResult = PyObject_CallMethod(dest, "write", "y#", output.dst,
+                                              output.pos);
+            if (NULL == writeResult) {
+                res = NULL;
+                goto finally;
+            }
+            totalWrite += output.pos;
+            Py_XDECREF(writeResult);
+            output.pos = 0;
+        }
+
+        if (!zresult) {
+            break;
+        }
+    }
+
+    totalReadPy = PyLong_FromSsize_t(totalRead);
+    totalWritePy = PyLong_FromSsize_t(totalWrite);
+    res = PyTuple_Pack(2, totalReadPy, totalWritePy);
+    Py_DECREF(totalReadPy);
+    Py_DECREF(totalWritePy);
+
+finally:
+    if (output.dst) {
+        PyMem_Free(output.dst);
+    }
+
+    Py_XDECREF(readResult);
+
+    return res;
+}
+
+static ZstdCompressionReader *ZstdCompressor_stream_reader(ZstdCompressor *self,
+                                                           PyObject *args,
+                                                           PyObject *kwargs) {
+    static char *kwlist[] = {"source", "size", "read_size", "closefd", NULL};
+
+    PyObject *source;
+    unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    size_t readSize = ZSTD_CStreamInSize();
+    PyObject *closefd = NULL;
+    ZstdCompressionReader *result = NULL;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|KkO:stream_reader",
+                                     kwlist, &source, &sourceSize, &readSize,
+                                     &closefd)) {
+        return NULL;
+    }
+
+    result = (ZstdCompressionReader *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionReaderType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    result->entered = 0;
+    result->closed = 0;
+
+    if (PyObject_HasAttrString(source, "read")) {
+        result->reader = source;
+        Py_INCREF(source);
+        result->readSize = readSize;
+    }
+    else if (1 == PyObject_CheckBuffer(source)) {
+        if (0 != PyObject_GetBuffer(source, &result->buffer, PyBUF_CONTIG_RO)) {
+            goto except;
+        }
+
+        assert(result->buffer.len >= 0);
+
+        sourceSize = result->buffer.len;
+    }
+    else {
+        PyErr_SetString(PyExc_TypeError,
+                        "must pass an object with a read() method or that "
+                        "conforms to the buffer protocol");
+        goto except;
+    }
+
+    result->closefd = closefd ? PyObject_IsTrue(closefd) : 1;
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source source: %s",
+                     ZSTD_getErrorName(zresult));
+        goto except;
+    }
+
+    result->compressor = self;
+    Py_INCREF(self);
+
+    return result;
+
+except:
+    Py_CLEAR(result);
+
+    return NULL;
+}
+
+static PyObject *ZstdCompressor_compress(ZstdCompressor *self, PyObject *args,
+                                         PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    Py_buffer source;
+    size_t destSize;
+    PyObject *output = NULL;
+    size_t zresult;
+    ZSTD_outBuffer outBuffer;
+    ZSTD_inBuffer inBuffer;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|O:compress", kwlist,
+                                     &source)) {
+        return NULL;
+    }
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    destSize = ZSTD_compressBound(source.len);
+    output = PyBytes_FromStringAndSize(NULL, destSize);
+    if (!output) {
+        goto finally;
+    }
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, source.len);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source size: %s",
+                     ZSTD_getErrorName(zresult));
+        Py_CLEAR(output);
+        goto finally;
+    }
+
+    inBuffer.src = source.buf;
+    inBuffer.size = source.len;
+    inBuffer.pos = 0;
+
+    outBuffer.dst = PyBytes_AsString(output);
+    outBuffer.size = destSize;
+    outBuffer.pos = 0;
+
+    Py_BEGIN_ALLOW_THREADS
+        /* By avoiding ZSTD_compress(), we don't necessarily write out content
+                size. This means the argument to ZstdCompressor to control frame
+                parameters is honored. */
+        zresult =
+            ZSTD_compressStream2(self->cctx, &outBuffer, &inBuffer, ZSTD_e_end);
+    Py_END_ALLOW_THREADS
+
+        if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "cannot compress: %s",
+                     ZSTD_getErrorName(zresult));
+        Py_CLEAR(output);
+        goto finally;
+    }
+    else if (zresult) {
+        PyErr_SetString(ZstdError, "unexpected partial frame flush");
+        Py_CLEAR(output);
+        goto finally;
+    }
+
+    Py_SET_SIZE(output, outBuffer.pos);
+
+finally:
+    PyBuffer_Release(&source);
+    return output;
+}
+
+static ZstdCompressionObj *ZstdCompressor_compressobj(ZstdCompressor *self,
+                                                      PyObject *args,
+                                                      PyObject *kwargs) {
+    static char *kwlist[] = {"size", NULL};
+
+    unsigned long long inSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    size_t outSize = ZSTD_CStreamOutSize();
+    ZstdCompressionObj *result = NULL;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|K:compressobj", kwlist,
+                                     &inSize)) {
+        return NULL;
+    }
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, inSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source size: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    result = (ZstdCompressionObj *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionObjType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    result->output.dst = PyMem_Malloc(outSize);
+    if (!result->output.dst) {
+        PyErr_NoMemory();
+        Py_DECREF(result);
+        return NULL;
+    }
+    result->output.size = outSize;
+    result->compressor = self;
+    Py_INCREF(result->compressor);
+
+    return result;
+}
+
+static ZstdCompressorIterator *ZstdCompressor_read_to_iter(ZstdCompressor *self,
+                                                           PyObject *args,
+                                                           PyObject *kwargs) {
+    static char *kwlist[] = {"reader", "size", "read_size", "write_size", NULL};
+
+    PyObject *reader;
+    unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    size_t inSize = ZSTD_CStreamInSize();
+    size_t outSize = ZSTD_CStreamOutSize();
+    ZstdCompressorIterator *result;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|Kkk:read_to_iter", kwlist,
+                                     &reader, &sourceSize, &inSize, &outSize)) {
+        return NULL;
+    }
+
+    result = (ZstdCompressorIterator *)PyObject_CallObject(
+        (PyObject *)ZstdCompressorIteratorType, NULL);
+    if (!result) {
+        return NULL;
+    }
+    if (PyObject_HasAttrString(reader, "read")) {
+        result->reader = reader;
+        Py_INCREF(result->reader);
+    }
+    else if (1 == PyObject_CheckBuffer(reader)) {
+        if (0 != PyObject_GetBuffer(reader, &result->buffer, PyBUF_CONTIG_RO)) {
+            goto except;
+        }
+
+        sourceSize = result->buffer.len;
+    }
+    else {
+        PyErr_SetString(PyExc_ValueError,
+                        "must pass an object with a read() method or conforms "
+                        "to buffer protocol");
+        goto except;
+    }
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source size: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    result->compressor = self;
+    Py_INCREF(result->compressor);
+
+    result->inSize = inSize;
+    result->outSize = outSize;
+
+    result->output.dst = PyMem_Malloc(outSize);
+    if (!result->output.dst) {
+        PyErr_NoMemory();
+        goto except;
+    }
+    result->output.size = outSize;
+
+    goto finally;
+
+except:
+    Py_CLEAR(result);
+
+finally:
+    return result;
+}
+
+static ZstdCompressionWriter *ZstdCompressor_stream_writer(ZstdCompressor *self,
+                                                           PyObject *args,
+                                                           PyObject *kwargs) {
+    static char *kwlist[] = {
+        "writer", "size", "write_size", "write_return_read", "closefd", NULL};
+
+    PyObject *writer;
+    ZstdCompressionWriter *result;
+    size_t zresult;
+    unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    size_t outSize = ZSTD_CStreamOutSize();
+    PyObject *writeReturnRead = NULL;
+    PyObject *closefd = NULL;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|KkOO:stream_writer",
+                                     kwlist, &writer, &sourceSize, &outSize,
+                                     &writeReturnRead, &closefd)) {
+        return NULL;
+    }
+
+    if (!PyObject_HasAttrString(writer, "write")) {
+        PyErr_SetString(PyExc_ValueError,
+                        "must pass an object with a write() method");
+        return NULL;
+    }
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source size: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    result = (ZstdCompressionWriter *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionWriterType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    result->entered = 0;
+    result->closing = 0;
+    result->closed = 0;
+
+    result->output.dst = PyMem_Malloc(outSize);
+    if (!result->output.dst) {
+        Py_DECREF(result);
+        return (ZstdCompressionWriter *)PyErr_NoMemory();
+    }
+
+    result->output.pos = 0;
+    result->output.size = outSize;
+
+    result->compressor = self;
+    Py_INCREF(result->compressor);
+
+    result->writer = writer;
+    Py_INCREF(result->writer);
+
+    result->outSize = outSize;
+    result->bytesCompressed = 0;
+    result->writeReturnRead =
+        writeReturnRead ? PyObject_IsTrue(writeReturnRead) : 1;
+    result->closefd = closefd ? PyObject_IsTrue(closefd) : 1;
+
+    return result;
+}
+
+static ZstdCompressionChunker *
+ZstdCompressor_chunker(ZstdCompressor *self, PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"size", "chunk_size", NULL};
+
+    unsigned long long sourceSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    size_t chunkSize = ZSTD_CStreamOutSize();
+    ZstdCompressionChunker *chunker;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|Kk:chunker", kwlist,
+                                     &sourceSize, &chunkSize)) {
+        return NULL;
+    }
+
+    ZSTD_CCtx_reset(self->cctx, ZSTD_reset_session_only);
+
+    zresult = ZSTD_CCtx_setPledgedSrcSize(self->cctx, sourceSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "error setting source size: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    chunker = (ZstdCompressionChunker *)PyObject_CallObject(
+        (PyObject *)ZstdCompressionChunkerType, NULL);
+    if (!chunker) {
+        return NULL;
+    }
+
+    chunker->output.dst = PyMem_Malloc(chunkSize);
+    if (!chunker->output.dst) {
+        PyErr_NoMemory();
+        Py_DECREF(chunker);
+        return NULL;
+    }
+    chunker->output.size = chunkSize;
+    chunker->output.pos = 0;
+
+    chunker->compressor = self;
+    Py_INCREF(chunker->compressor);
+
+    chunker->chunkSize = chunkSize;
+
+    return chunker;
+}
+
+typedef struct {
+    void *sourceData;
+    size_t sourceSize;
+} DataSource;
+
+typedef struct {
+    DataSource *sources;
+    Py_ssize_t sourcesSize;
+    unsigned long long totalSourceSize;
+} DataSources;
+
+typedef struct {
+    void *dest;
+    Py_ssize_t destSize;
+    BufferSegment *segments;
+    Py_ssize_t segmentsSize;
+} CompressorDestBuffer;
+
+typedef enum {
+    CompressorWorkerError_none = 0,
+    CompressorWorkerError_zstd = 1,
+    CompressorWorkerError_no_memory = 2,
+    CompressorWorkerError_nospace = 3,
+} CompressorWorkerError;
+
+/**
+ * Holds state for an individual worker performing multi_compress_to_buffer
+ * work.
+ */
+typedef struct {
+    /* Used for compression. */
+    ZSTD_CCtx *cctx;
+
+    /* What to compress. */
+    DataSource *sources;
+    Py_ssize_t sourcesSize;
+    Py_ssize_t startOffset;
+    Py_ssize_t endOffset;
+    unsigned long long totalSourceSize;
+
+    /* Result storage. */
+    CompressorDestBuffer *destBuffers;
+    Py_ssize_t destCount;
+
+    /* Error tracking. */
+    CompressorWorkerError error;
+    size_t zresult;
+    Py_ssize_t errorOffset;
+} CompressorWorkerState;
+
+#ifdef HAVE_ZSTD_POOL_APIS
+static void compress_worker(CompressorWorkerState *state) {
+    Py_ssize_t inputOffset = state->startOffset;
+    Py_ssize_t remainingItems = state->endOffset - state->startOffset + 1;
+    Py_ssize_t currentBufferStartOffset = state->startOffset;
+    size_t zresult;
+    void *newDest;
+    size_t allocationSize;
+    size_t boundSize;
+    Py_ssize_t destOffset = 0;
+    DataSource *sources = state->sources;
+    CompressorDestBuffer *destBuffer;
+
+    assert(!state->destBuffers);
+    assert(0 == state->destCount);
+
+    /*
+     * The total size of the compressed data is unknown until we actually
+     * compress data. That means we can't pre-allocate the exact size we need.
+     *
+     * There is a cost to every allocation and reallocation. So, it is in our
+     * interest to minimize the number of allocations.
+     *
+     * There is also a cost to too few allocations. If allocations are too
+     * large they may fail. If buffers are shared and all inputs become
+     * irrelevant at different lifetimes, then a reference to one segment
+     * in the buffer will keep the entire buffer alive. This leads to excessive
+     * memory usage.
+     *
+     * Our current strategy is to assume a compression ratio of 16:1 and
+     * allocate buffers of that size, rounded up to the nearest power of 2
+     * (because computers like round numbers). That ratio is greater than what
+     * most inputs achieve. This is by design: we don't want to over-allocate.
+     * But we don't want to under-allocate and lead to too many buffers either.
+     */
+
+    state->destCount = 1;
+
+    state->destBuffers = calloc(1, sizeof(CompressorDestBuffer));
+    if (NULL == state->destBuffers) {
+        state->error = CompressorWorkerError_no_memory;
+        return;
+    }
+
+    destBuffer = &state->destBuffers[state->destCount - 1];
+
+    /*
+     * Rather than track bounds and grow the segments buffer, allocate space
+     * to hold remaining items then truncate when we're done with it.
+     */
+    destBuffer->segments = calloc(remainingItems, sizeof(BufferSegment));
+    if (NULL == destBuffer->segments) {
+        state->error = CompressorWorkerError_no_memory;
+        return;
+    }
+
+    destBuffer->segmentsSize = remainingItems;
+
+    assert(state->totalSourceSize <= SIZE_MAX);
+    allocationSize = roundpow2((size_t)state->totalSourceSize >> 4);
+
+    /* If the maximum size of the output is larger than that, round up. */
+    boundSize = ZSTD_compressBound(sources[inputOffset].sourceSize);
+
+    if (boundSize > allocationSize) {
+        allocationSize = roundpow2(boundSize);
+    }
+
+    destBuffer->dest = malloc(allocationSize);
+    if (NULL == destBuffer->dest) {
+        state->error = CompressorWorkerError_no_memory;
+        return;
+    }
+
+    destBuffer->destSize = allocationSize;
+
+    for (inputOffset = state->startOffset; inputOffset <= state->endOffset;
+         inputOffset++) {
+        void *source = sources[inputOffset].sourceData;
+        size_t sourceSize = sources[inputOffset].sourceSize;
+        size_t destAvailable;
+        void *dest;
+        ZSTD_outBuffer opOutBuffer;
+        ZSTD_inBuffer opInBuffer;
+
+        destAvailable = destBuffer->destSize - destOffset;
+        boundSize = ZSTD_compressBound(sourceSize);
+
+        /*
+         * Not enough space in current buffer to hold largest compressed output.
+         * So allocate and switch to a new output buffer.
+         */
+        if (boundSize > destAvailable) {
+            /*
+             * The downsizing of the existing buffer is optional. It should be
+             * cheap (unlike growing). So we just do it.
+             */
+            if (destAvailable) {
+                newDest = realloc(destBuffer->dest, destOffset);
+                if (NULL == newDest) {
+                    state->error = CompressorWorkerError_no_memory;
+                    return;
+                }
+
+                destBuffer->dest = newDest;
+                destBuffer->destSize = destOffset;
+            }
+
+            /* Truncate segments buffer. */
+            newDest = realloc(destBuffer->segments,
+                              (inputOffset - currentBufferStartOffset + 1) *
+                                  sizeof(BufferSegment));
+            if (NULL == newDest) {
+                state->error = CompressorWorkerError_no_memory;
+                return;
+            }
+
+            destBuffer->segments = newDest;
+            destBuffer->segmentsSize = inputOffset - currentBufferStartOffset;
+
+            /* Grow space for new struct. */
+            /* TODO consider over-allocating so we don't do this every time. */
+            newDest =
+                realloc(state->destBuffers,
+                        (state->destCount + 1) * sizeof(CompressorDestBuffer));
+            if (NULL == newDest) {
+                state->error = CompressorWorkerError_no_memory;
+                return;
+            }
+
+            state->destBuffers = newDest;
+            state->destCount++;
+
+            destBuffer = &state->destBuffers[state->destCount - 1];
+
+            /* Don't take any chances with non-NULL pointers. */
+            memset(destBuffer, 0, sizeof(CompressorDestBuffer));
+
+            /**
+             * We could dynamically update allocation size based on work done so
+             * far. For now, keep is simple.
+             */
+            assert(state->totalSourceSize <= SIZE_MAX);
+            allocationSize = roundpow2((size_t)state->totalSourceSize >> 4);
+
+            if (boundSize > allocationSize) {
+                allocationSize = roundpow2(boundSize);
+            }
+
+            destBuffer->dest = malloc(allocationSize);
+            if (NULL == destBuffer->dest) {
+                state->error = CompressorWorkerError_no_memory;
+                return;
+            }
+
+            destBuffer->destSize = allocationSize;
+            destAvailable = allocationSize;
+            destOffset = 0;
+
+            destBuffer->segments =
+                calloc(remainingItems, sizeof(BufferSegment));
+            if (NULL == destBuffer->segments) {
+                state->error = CompressorWorkerError_no_memory;
+                return;
+            }
+
+            destBuffer->segmentsSize = remainingItems;
+            currentBufferStartOffset = inputOffset;
+        }
+
+        dest = (char *)destBuffer->dest + destOffset;
+
+        opInBuffer.src = source;
+        opInBuffer.size = sourceSize;
+        opInBuffer.pos = 0;
+
+        opOutBuffer.dst = dest;
+        opOutBuffer.size = destAvailable;
+        opOutBuffer.pos = 0;
+
+        zresult = ZSTD_CCtx_setPledgedSrcSize(state->cctx, sourceSize);
+        if (ZSTD_isError(zresult)) {
+            state->error = CompressorWorkerError_zstd;
+            state->zresult = zresult;
+            state->errorOffset = inputOffset;
+            break;
+        }
+
+        zresult = ZSTD_compressStream2(state->cctx, &opOutBuffer, &opInBuffer,
+                                       ZSTD_e_end);
+        if (ZSTD_isError(zresult)) {
+            state->error = CompressorWorkerError_zstd;
+            state->zresult = zresult;
+            state->errorOffset = inputOffset;
+            break;
+        }
+        else if (zresult) {
+            state->error = CompressorWorkerError_nospace;
+            state->errorOffset = inputOffset;
+            break;
+        }
+
+        destBuffer->segments[inputOffset - currentBufferStartOffset].offset =
+            destOffset;
+        destBuffer->segments[inputOffset - currentBufferStartOffset].length =
+            opOutBuffer.pos;
+
+        destOffset += opOutBuffer.pos;
+        remainingItems--;
+    }
+
+    if (destBuffer->destSize > destOffset) {
+        newDest = realloc(destBuffer->dest, destOffset);
+        if (NULL == newDest) {
+            state->error = CompressorWorkerError_no_memory;
+            return;
+        }
+
+        destBuffer->dest = newDest;
+        destBuffer->destSize = destOffset;
+    }
+}
+#endif
+
+/* We can only use the pool.h APIs if we provide the full library,
+   as these are private APIs. */
+#ifdef HAVE_ZSTD_POOL_APIS
+
+ZstdBufferWithSegmentsCollection *
+compress_from_datasources(ZstdCompressor *compressor, DataSources *sources,
+                          Py_ssize_t threadCount) {
+    unsigned long long bytesPerWorker;
+    POOL_ctx *pool = NULL;
+    CompressorWorkerState *workerStates = NULL;
+    Py_ssize_t i;
+    unsigned long long workerBytes = 0;
+    Py_ssize_t workerStartOffset = 0;
+    Py_ssize_t currentThread = 0;
+    int errored = 0;
+    Py_ssize_t segmentsCount = 0;
+    Py_ssize_t segmentIndex;
+    PyObject *segmentsArg = NULL;
+    ZstdBufferWithSegments *buffer;
+    ZstdBufferWithSegmentsCollection *result = NULL;
+
+    assert(sources->sourcesSize > 0);
+    assert(sources->totalSourceSize > 0);
+    assert(threadCount >= 1);
+
+    /* More threads than inputs makes no sense. */
+    threadCount =
+        sources->sourcesSize < threadCount ? sources->sourcesSize : threadCount;
+
+    /* TODO lower thread count when input size is too small and threads would
+    add overhead. */
+
+    workerStates = PyMem_Malloc(threadCount * sizeof(CompressorWorkerState));
+    if (NULL == workerStates) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+
+    memset(workerStates, 0, threadCount * sizeof(CompressorWorkerState));
+
+    if (threadCount > 1) {
+        pool = POOL_create(threadCount, 1);
+        if (NULL == pool) {
+            PyErr_SetString(ZstdError, "could not initialize zstd thread pool");
+            goto finally;
+        }
+    }
+
+    bytesPerWorker = sources->totalSourceSize / threadCount;
+
+    for (i = 0; i < threadCount; i++) {
+        size_t zresult;
+
+        workerStates[i].cctx = ZSTD_createCCtx();
+        if (!workerStates[i].cctx) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        zresult = ZSTD_CCtx_setParametersUsingCCtxParams(workerStates[i].cctx,
+                                                         compressor->params);
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "could not set compression parameters: %s",
+                         ZSTD_getErrorName(zresult));
+            goto finally;
+        }
+
+        if (compressor->dict) {
+            if (compressor->dict->cdict) {
+                zresult = ZSTD_CCtx_refCDict(workerStates[i].cctx,
+                                             compressor->dict->cdict);
+            }
+            else {
+                zresult = ZSTD_CCtx_loadDictionary_advanced(
+                    workerStates[i].cctx, compressor->dict->dictData,
+                    compressor->dict->dictSize, ZSTD_dlm_byRef,
+                    compressor->dict->dictType);
+            }
+
+            if (ZSTD_isError(zresult)) {
+                PyErr_Format(ZstdError,
+                             "could not load compression dictionary: %s",
+                             ZSTD_getErrorName(zresult));
+                goto finally;
+            }
+        }
+
+        workerStates[i].sources = sources->sources;
+        workerStates[i].sourcesSize = sources->sourcesSize;
+    }
+
+    Py_BEGIN_ALLOW_THREADS for (i = 0; i < sources->sourcesSize; i++) {
+        workerBytes += sources->sources[i].sourceSize;
+
+        /*
+         * The last worker/thread needs to handle all remaining work. Don't
+         * trigger it prematurely. Defer to the block outside of the loop
+         * to run the last worker/thread. But do still process this loop
+         * so workerBytes is correct.
+         */
+        if (currentThread == threadCount - 1) {
+            continue;
+        }
+
+        if (workerBytes >= bytesPerWorker) {
+            assert(currentThread < threadCount);
+            workerStates[currentThread].totalSourceSize = workerBytes;
+            workerStates[currentThread].startOffset = workerStartOffset;
+            workerStates[currentThread].endOffset = i;
+
+            if (threadCount > 1) {
+                POOL_add(pool, (POOL_function)compress_worker,
+                         &workerStates[currentThread]);
+            }
+            else {
+                compress_worker(&workerStates[currentThread]);
+            }
+
+            currentThread++;
+            workerStartOffset = i + 1;
+            workerBytes = 0;
+        }
+    }
+
+    if (workerBytes) {
+        assert(currentThread < threadCount);
+        workerStates[currentThread].totalSourceSize = workerBytes;
+        workerStates[currentThread].startOffset = workerStartOffset;
+        workerStates[currentThread].endOffset = sources->sourcesSize - 1;
+
+        if (threadCount > 1) {
+            POOL_add(pool, (POOL_function)compress_worker,
+                     &workerStates[currentThread]);
+        }
+        else {
+            compress_worker(&workerStates[currentThread]);
+        }
+    }
+
+    if (threadCount > 1) {
+        POOL_free(pool);
+        pool = NULL;
+    }
+
+    Py_END_ALLOW_THREADS
+
+        for (i = 0; i < threadCount; i++) {
+        switch (workerStates[i].error) {
+        case CompressorWorkerError_no_memory:
+            PyErr_NoMemory();
+            errored = 1;
+            break;
+
+        case CompressorWorkerError_zstd:
+            PyErr_Format(ZstdError, "error compressing item %zd: %s",
+                         workerStates[i].errorOffset,
+                         ZSTD_getErrorName(workerStates[i].zresult));
+            errored = 1;
+            break;
+
+        case CompressorWorkerError_nospace:
+            PyErr_Format(
+                ZstdError,
+                "error compressing item %zd: not enough space in output",
+                workerStates[i].errorOffset);
+            errored = 1;
+            break;
+
+        default:;
+        }
+
+        if (errored) {
+            break;
+        }
+    }
+
+    if (errored) {
+        goto finally;
+    }
+
+    segmentsCount = 0;
+    for (i = 0; i < threadCount; i++) {
+        CompressorWorkerState *state = &workerStates[i];
+        segmentsCount += state->destCount;
+    }
+
+    segmentsArg = PyTuple_New(segmentsCount);
+    if (NULL == segmentsArg) {
+        goto finally;
+    }
+
+    segmentIndex = 0;
+
+    for (i = 0; i < threadCount; i++) {
+        Py_ssize_t j;
+        CompressorWorkerState *state = &workerStates[i];
+
+        for (j = 0; j < state->destCount; j++) {
+            CompressorDestBuffer *destBuffer = &state->destBuffers[j];
+            buffer = BufferWithSegments_FromMemory(
+                destBuffer->dest, destBuffer->destSize, destBuffer->segments,
+                destBuffer->segmentsSize);
+
+            if (NULL == buffer) {
+                goto finally;
+            }
+
+            /* Tell instance to use free() instsead of PyMem_Free(). */
+            buffer->useFree = 1;
+
+            /*
+             * BufferWithSegments_FromMemory takes ownership of the backing
+             * memory. Unset it here so it doesn't get freed below.
+             */
+            destBuffer->dest = NULL;
+            destBuffer->segments = NULL;
+
+            PyTuple_SET_ITEM(segmentsArg, segmentIndex++, (PyObject *)buffer);
+        }
+    }
+
+    result = (ZstdBufferWithSegmentsCollection *)PyObject_CallObject(
+        (PyObject *)ZstdBufferWithSegmentsCollectionType, segmentsArg);
+
+finally:
+    Py_CLEAR(segmentsArg);
+
+    if (pool) {
+        POOL_free(pool);
+    }
+
+    if (workerStates) {
+        Py_ssize_t j;
+
+        for (i = 0; i < threadCount; i++) {
+            CompressorWorkerState state = workerStates[i];
+
+            if (state.cctx) {
+                ZSTD_freeCCtx(state.cctx);
+            }
+
+            /* malloc() is used in worker thread. */
+
+            for (j = 0; j < state.destCount; j++) {
+                if (state.destBuffers) {
+                    free(state.destBuffers[j].dest);
+                    free(state.destBuffers[j].segments);
+                }
+            }
+
+            free(state.destBuffers);
+        }
+
+        PyMem_Free(workerStates);
+    }
+
+    return result;
+}
+#endif
+
+#ifdef HAVE_ZSTD_POOL_APIS
+static ZstdBufferWithSegmentsCollection *
+ZstdCompressor_multi_compress_to_buffer(ZstdCompressor *self, PyObject *args,
+                                        PyObject *kwargs) {
+    static char *kwlist[] = {"data", "threads", NULL};
+
+    PyObject *data;
+    int threads = 0;
+    Py_buffer *dataBuffers = NULL;
+    DataSources sources;
+    Py_ssize_t i;
+    Py_ssize_t sourceCount = 0;
+    ZstdBufferWithSegmentsCollection *result = NULL;
+
+    memset(&sources, 0, sizeof(sources));
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs,
+                                     "O|i:multi_compress_to_buffer", kwlist,
+                                     &data, &threads)) {
+        return NULL;
+    }
+
+    if (threads < 0) {
+        threads = cpu_count();
+    }
+
+    if (threads < 2) {
+        threads = 1;
+    }
+
+    if (PyObject_TypeCheck(data, ZstdBufferWithSegmentsType)) {
+        ZstdBufferWithSegments *buffer = (ZstdBufferWithSegments *)data;
+
+        sources.sources =
+            PyMem_Malloc(buffer->segmentCount * sizeof(DataSource));
+        if (NULL == sources.sources) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        for (i = 0; i < buffer->segmentCount; i++) {
+            if (buffer->segments[i].length > SIZE_MAX) {
+                PyErr_Format(
+                    PyExc_ValueError,
+                    "buffer segment %zd is too large for this platform", i);
+                goto finally;
+            }
+
+            sources.sources[i].sourceData =
+                (char *)buffer->data + buffer->segments[i].offset;
+            sources.sources[i].sourceSize = (size_t)buffer->segments[i].length;
+            sources.totalSourceSize += buffer->segments[i].length;
+        }
+
+        sources.sourcesSize = buffer->segmentCount;
+    }
+    else if (PyObject_TypeCheck(data, ZstdBufferWithSegmentsCollectionType)) {
+        Py_ssize_t j;
+        Py_ssize_t offset = 0;
+        ZstdBufferWithSegments *buffer;
+        ZstdBufferWithSegmentsCollection *collection =
+            (ZstdBufferWithSegmentsCollection *)data;
+
+        sourceCount = BufferWithSegmentsCollection_length(collection);
+
+        sources.sources = PyMem_Malloc(sourceCount * sizeof(DataSource));
+        if (NULL == sources.sources) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        for (i = 0; i < collection->bufferCount; i++) {
+            buffer = collection->buffers[i];
+
+            for (j = 0; j < buffer->segmentCount; j++) {
+                if (buffer->segments[j].length > SIZE_MAX) {
+                    PyErr_Format(PyExc_ValueError,
+                                 "buffer segment %zd in buffer %zd is too "
+                                 "large for this platform",
+                                 j, i);
+                    goto finally;
+                }
+
+                sources.sources[offset].sourceData =
+                    (char *)buffer->data + buffer->segments[j].offset;
+                sources.sources[offset].sourceSize =
+                    (size_t)buffer->segments[j].length;
+                sources.totalSourceSize += buffer->segments[j].length;
+
+                offset++;
+            }
+        }
+
+        sources.sourcesSize = sourceCount;
+    }
+    else if (PyList_Check(data)) {
+        sourceCount = PyList_GET_SIZE(data);
+
+        sources.sources = PyMem_Malloc(sourceCount * sizeof(DataSource));
+        if (NULL == sources.sources) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        dataBuffers = PyMem_Malloc(sourceCount * sizeof(Py_buffer));
+        if (NULL == dataBuffers) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        memset(dataBuffers, 0, sourceCount * sizeof(Py_buffer));
+
+        for (i = 0; i < sourceCount; i++) {
+            if (0 != PyObject_GetBuffer(PyList_GET_ITEM(data, i),
+                                        &dataBuffers[i], PyBUF_CONTIG_RO)) {
+                PyErr_Clear();
+                PyErr_Format(PyExc_TypeError,
+                             "item %zd not a bytes like object", i);
+                goto finally;
+            }
+
+            sources.sources[i].sourceData = dataBuffers[i].buf;
+            sources.sources[i].sourceSize = dataBuffers[i].len;
+            sources.totalSourceSize += dataBuffers[i].len;
+        }
+
+        sources.sourcesSize = sourceCount;
+    }
+    else {
+        PyErr_SetString(PyExc_TypeError,
+                        "argument must be list of BufferWithSegments");
+        goto finally;
+    }
+
+    if (0 == sources.sourcesSize) {
+        PyErr_SetString(PyExc_ValueError, "no source elements found");
+        goto finally;
+    }
+
+    if (0 == sources.totalSourceSize) {
+        PyErr_SetString(PyExc_ValueError, "source elements are empty");
+        goto finally;
+    }
+
+    if (sources.totalSourceSize > SIZE_MAX) {
+        PyErr_SetString(PyExc_ValueError,
+                        "sources are too large for this platform");
+        goto finally;
+    }
+
+    result = compress_from_datasources(self, &sources, threads);
+
+finally:
+    PyMem_Free(sources.sources);
+
+    if (dataBuffers) {
+        for (i = 0; i < sourceCount; i++) {
+            PyBuffer_Release(&dataBuffers[i]);
+        }
+
+        PyMem_Free(dataBuffers);
+    }
+
+    return result;
+}
+#endif
+
+static PyMethodDef ZstdCompressor_methods[] = {
+    {"chunker", (PyCFunction)ZstdCompressor_chunker,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"compress", (PyCFunction)ZstdCompressor_compress,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"compressobj", (PyCFunction)ZstdCompressor_compressobj,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"copy_stream", (PyCFunction)ZstdCompressor_copy_stream,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"stream_reader", (PyCFunction)ZstdCompressor_stream_reader,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"stream_writer", (PyCFunction)ZstdCompressor_stream_writer,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"read_to_iter", (PyCFunction)ZstdCompressor_read_to_iter,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+#ifdef HAVE_ZSTD_POOL_APIS
+    {"multi_compress_to_buffer",
+     (PyCFunction)ZstdCompressor_multi_compress_to_buffer,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+#endif
+    {"memory_size", (PyCFunction)ZstdCompressor_memory_size, METH_NOARGS, NULL},
+    {"frame_progression", (PyCFunction)ZstdCompressor_frame_progression,
+     METH_NOARGS, NULL},
+    {NULL, NULL}};
+
+PyType_Slot ZstdCompressorSlots[] = {
+    {Py_tp_dealloc, ZstdCompressor_dealloc},
+    {Py_tp_methods, ZstdCompressor_methods},
+    {Py_tp_init, ZstdCompressor_init},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressorSpec = {
+    "zstd.ZstdCompressor",
+    sizeof(ZstdCompressor),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressorSlots,
+};
+
+PyTypeObject *ZstdCompressorType;
+
+void compressor_module_init(PyObject *mod) {
+    ZstdCompressorType = (PyTypeObject *)PyType_FromSpec(&ZstdCompressorSpec);
+    if (PyType_Ready(ZstdCompressorType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdCompressorType);
+    PyModule_AddObject(mod, "ZstdCompressor", (PyObject *)ZstdCompressorType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/compressoriterator.c b/contrib/python/zstandard/py3/c-ext/compressoriterator.c
new file mode 100644
index 00000000000..9e4936caa8b
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/compressoriterator.c
@@ -0,0 +1,212 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+#define min(a, b) (((a) < (b)) ? (a) : (b))
+
+extern PyObject *ZstdError;
+
+static void ZstdCompressorIterator_dealloc(ZstdCompressorIterator *self) {
+    Py_XDECREF(self->readResult);
+    Py_XDECREF(self->compressor);
+    Py_XDECREF(self->reader);
+
+    if (self->buffer.buf) {
+        PyBuffer_Release(&self->buffer);
+        memset(&self->buffer, 0, sizeof(self->buffer));
+    }
+
+    if (self->output.dst) {
+        PyMem_Free(self->output.dst);
+        self->output.dst = NULL;
+    }
+
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdCompressorIterator_iter(PyObject *self) {
+    Py_INCREF(self);
+    return self;
+}
+
+static PyObject *ZstdCompressorIterator_iternext(ZstdCompressorIterator *self) {
+    size_t zresult;
+    PyObject *readResult = NULL;
+    PyObject *chunk;
+    char *readBuffer;
+    Py_ssize_t readSize = 0;
+    Py_ssize_t bufferRemaining;
+
+    if (self->finishedOutput) {
+        PyErr_SetString(PyExc_StopIteration, "output flushed");
+        return NULL;
+    }
+
+feedcompressor:
+
+    /* If we have data left in the input, consume it. */
+    if (self->input.pos < self->input.size) {
+        Py_BEGIN_ALLOW_THREADS zresult =
+            ZSTD_compressStream2(self->compressor->cctx, &self->output,
+                                 &self->input, ZSTD_e_continue);
+        Py_END_ALLOW_THREADS
+
+            /* Release the Python object holding the input buffer. */
+            if (self->input.pos == self->input.size) {
+            self->input.src = NULL;
+            self->input.pos = 0;
+            self->input.size = 0;
+            Py_DECREF(self->readResult);
+            self->readResult = NULL;
+        }
+
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd compress error: %s",
+                         ZSTD_getErrorName(zresult));
+            return NULL;
+        }
+
+        /* If it produced output data, emit it. */
+        if (self->output.pos) {
+            chunk =
+                PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+            self->output.pos = 0;
+            return chunk;
+        }
+    }
+
+    /* We should never have output data sitting around after a previous call. */
+    assert(self->output.pos == 0);
+
+    /* The code above should have either emitted a chunk and returned or
+    consumed the entire input buffer. So the state of the input buffer is not
+    relevant. */
+    if (!self->finishedInput) {
+        if (self->reader) {
+            readResult =
+                PyObject_CallMethod(self->reader, "read", "I", self->inSize);
+            if (!readResult) {
+                return NULL;
+            }
+
+            PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+        }
+        else {
+            assert(self->buffer.buf);
+
+            /* Only support contiguous C arrays. */
+            assert(self->buffer.strides == NULL &&
+                   self->buffer.suboffsets == NULL);
+            assert(self->buffer.itemsize == 1);
+
+            readBuffer = (char *)self->buffer.buf + self->bufferOffset;
+            bufferRemaining = self->buffer.len - self->bufferOffset;
+            readSize = min(bufferRemaining, (Py_ssize_t)self->inSize);
+            self->bufferOffset += readSize;
+        }
+
+        if (0 == readSize) {
+            Py_XDECREF(readResult);
+            self->finishedInput = 1;
+        }
+        else {
+            self->readResult = readResult;
+        }
+    }
+
+    /* EOF */
+    if (0 == readSize) {
+        self->input.src = NULL;
+        self->input.size = 0;
+        self->input.pos = 0;
+
+        zresult = ZSTD_compressStream2(self->compressor->cctx, &self->output,
+                                       &self->input, ZSTD_e_end);
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "error ending compression stream: %s",
+                         ZSTD_getErrorName(zresult));
+            return NULL;
+        }
+
+        assert(self->output.pos);
+
+        if (0 == zresult) {
+            self->finishedOutput = 1;
+        }
+
+        chunk = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+        self->output.pos = 0;
+        return chunk;
+    }
+
+    /* New data from reader. Feed into compressor. */
+    self->input.src = readBuffer;
+    self->input.size = readSize;
+    self->input.pos = 0;
+
+    Py_BEGIN_ALLOW_THREADS zresult = ZSTD_compressStream2(
+        self->compressor->cctx, &self->output, &self->input, ZSTD_e_continue);
+    Py_END_ALLOW_THREADS
+
+        /* The input buffer currently points to memory managed by Python
+        (readBuffer). This object was allocated by this function. If it wasn't
+        fully consumed, we need to release it in a subsequent function call.
+        If it is fully consumed, do that now.
+        */
+        if (self->input.pos == self->input.size) {
+        self->input.src = NULL;
+        self->input.pos = 0;
+        self->input.size = 0;
+        Py_XDECREF(self->readResult);
+        self->readResult = NULL;
+    }
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "zstd compress error: %s",
+                     ZSTD_getErrorName(zresult));
+        return NULL;
+    }
+
+    assert(self->input.pos <= self->input.size);
+
+    /* If we didn't write anything, start the process over. */
+    if (0 == self->output.pos) {
+        goto feedcompressor;
+    }
+
+    chunk = PyBytes_FromStringAndSize(self->output.dst, self->output.pos);
+    self->output.pos = 0;
+    return chunk;
+}
+
+PyType_Slot ZstdCompressorIteratorSlots[] = {
+    {Py_tp_dealloc, ZstdCompressorIterator_dealloc},
+    {Py_tp_iter, ZstdCompressorIterator_iter},
+    {Py_tp_iternext, ZstdCompressorIterator_iternext},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdCompressorIteratorSpec = {
+    "zstd.ZstdCompressorIterator",
+    sizeof(ZstdCompressorIterator),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdCompressorIteratorSlots,
+};
+
+PyTypeObject *ZstdCompressorIteratorType;
+
+void compressoriterator_module_init(PyObject *mod) {
+    ZstdCompressorIteratorType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdCompressorIteratorSpec);
+    if (PyType_Ready(ZstdCompressorIteratorType) < 0) {
+        return;
+    }
+}
diff --git a/contrib/python/zstandard/py3/c-ext/constants.c b/contrib/python/zstandard/py3/c-ext/constants.c
new file mode 100644
index 00000000000..e209ecf9a84
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/constants.c
@@ -0,0 +1,109 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static char frame_header[] = {
+    '\x28',
+    '\xb5',
+    '\x2f',
+    '\xfd',
+};
+
+void constants_module_init(PyObject *mod) {
+    PyObject *version;
+    PyObject *zstdVersion;
+    PyObject *frameHeader;
+
+    version = PyUnicode_FromString(PYTHON_ZSTANDARD_VERSION);
+    PyModule_AddObject(mod, "__version__", version);
+
+    ZstdError = PyErr_NewException("zstd.ZstdError", NULL, NULL);
+    PyModule_AddObject(mod, "ZstdError", ZstdError);
+
+    PyModule_AddIntConstant(mod, "FLUSH_BLOCK", 0);
+    PyModule_AddIntConstant(mod, "FLUSH_FRAME", 1);
+
+    PyModule_AddIntConstant(mod, "COMPRESSOBJ_FLUSH_FINISH",
+                            compressorobj_flush_finish);
+    PyModule_AddIntConstant(mod, "COMPRESSOBJ_FLUSH_BLOCK",
+                            compressorobj_flush_block);
+
+    /* For now, the version is a simple tuple instead of a dedicated type. */
+    zstdVersion = PyTuple_New(3);
+    PyTuple_SetItem(zstdVersion, 0, PyLong_FromLong(ZSTD_VERSION_MAJOR));
+    PyTuple_SetItem(zstdVersion, 1, PyLong_FromLong(ZSTD_VERSION_MINOR));
+    PyTuple_SetItem(zstdVersion, 2, PyLong_FromLong(ZSTD_VERSION_RELEASE));
+    PyModule_AddObject(mod, "ZSTD_VERSION", zstdVersion);
+
+    frameHeader = PyBytes_FromStringAndSize(frame_header, sizeof(frame_header));
+    if (frameHeader) {
+        PyModule_AddObject(mod, "FRAME_HEADER", frameHeader);
+    }
+    else {
+        PyErr_Format(PyExc_ValueError, "could not create frame header object");
+    }
+
+    PyModule_AddObject(mod, "CONTENTSIZE_UNKNOWN",
+                       PyLong_FromUnsignedLongLong(ZSTD_CONTENTSIZE_UNKNOWN));
+    PyModule_AddObject(mod, "CONTENTSIZE_ERROR",
+                       PyLong_FromUnsignedLongLong(ZSTD_CONTENTSIZE_ERROR));
+
+    PyModule_AddIntConstant(mod, "MAX_COMPRESSION_LEVEL", ZSTD_maxCLevel());
+    PyModule_AddIntConstant(mod, "COMPRESSION_RECOMMENDED_INPUT_SIZE",
+                            (long)ZSTD_CStreamInSize());
+    PyModule_AddIntConstant(mod, "COMPRESSION_RECOMMENDED_OUTPUT_SIZE",
+                            (long)ZSTD_CStreamOutSize());
+    PyModule_AddIntConstant(mod, "DECOMPRESSION_RECOMMENDED_INPUT_SIZE",
+                            (long)ZSTD_DStreamInSize());
+    PyModule_AddIntConstant(mod, "DECOMPRESSION_RECOMMENDED_OUTPUT_SIZE",
+                            (long)ZSTD_DStreamOutSize());
+
+    PyModule_AddIntConstant(mod, "MAGIC_NUMBER", ZSTD_MAGICNUMBER);
+    PyModule_AddIntConstant(mod, "BLOCKSIZELOG_MAX", ZSTD_BLOCKSIZELOG_MAX);
+    PyModule_AddIntConstant(mod, "BLOCKSIZE_MAX", ZSTD_BLOCKSIZE_MAX);
+    PyModule_AddIntConstant(mod, "WINDOWLOG_MIN", ZSTD_WINDOWLOG_MIN);
+    PyModule_AddIntConstant(mod, "WINDOWLOG_MAX", ZSTD_WINDOWLOG_MAX);
+    PyModule_AddIntConstant(mod, "CHAINLOG_MIN", ZSTD_CHAINLOG_MIN);
+    PyModule_AddIntConstant(mod, "CHAINLOG_MAX", ZSTD_CHAINLOG_MAX);
+    PyModule_AddIntConstant(mod, "HASHLOG_MIN", ZSTD_HASHLOG_MIN);
+    PyModule_AddIntConstant(mod, "HASHLOG_MAX", ZSTD_HASHLOG_MAX);
+    PyModule_AddIntConstant(mod, "SEARCHLOG_MIN", ZSTD_SEARCHLOG_MIN);
+    PyModule_AddIntConstant(mod, "SEARCHLOG_MAX", ZSTD_SEARCHLOG_MAX);
+    PyModule_AddIntConstant(mod, "MINMATCH_MIN", ZSTD_MINMATCH_MIN);
+    PyModule_AddIntConstant(mod, "MINMATCH_MAX", ZSTD_MINMATCH_MAX);
+    /* TODO SEARCHLENGTH_* is deprecated. */
+    PyModule_AddIntConstant(mod, "SEARCHLENGTH_MIN", ZSTD_MINMATCH_MIN);
+    PyModule_AddIntConstant(mod, "SEARCHLENGTH_MAX", ZSTD_MINMATCH_MAX);
+    PyModule_AddIntConstant(mod, "TARGETLENGTH_MIN", ZSTD_TARGETLENGTH_MIN);
+    PyModule_AddIntConstant(mod, "TARGETLENGTH_MAX", ZSTD_TARGETLENGTH_MAX);
+    PyModule_AddIntConstant(mod, "LDM_MINMATCH_MIN", ZSTD_LDM_MINMATCH_MIN);
+    PyModule_AddIntConstant(mod, "LDM_MINMATCH_MAX", ZSTD_LDM_MINMATCH_MAX);
+    PyModule_AddIntConstant(mod, "LDM_BUCKETSIZELOG_MAX",
+                            ZSTD_LDM_BUCKETSIZELOG_MAX);
+
+    PyModule_AddIntConstant(mod, "STRATEGY_FAST", ZSTD_fast);
+    PyModule_AddIntConstant(mod, "STRATEGY_DFAST", ZSTD_dfast);
+    PyModule_AddIntConstant(mod, "STRATEGY_GREEDY", ZSTD_greedy);
+    PyModule_AddIntConstant(mod, "STRATEGY_LAZY", ZSTD_lazy);
+    PyModule_AddIntConstant(mod, "STRATEGY_LAZY2", ZSTD_lazy2);
+    PyModule_AddIntConstant(mod, "STRATEGY_BTLAZY2", ZSTD_btlazy2);
+    PyModule_AddIntConstant(mod, "STRATEGY_BTOPT", ZSTD_btopt);
+    PyModule_AddIntConstant(mod, "STRATEGY_BTULTRA", ZSTD_btultra);
+    PyModule_AddIntConstant(mod, "STRATEGY_BTULTRA2", ZSTD_btultra2);
+
+    PyModule_AddIntConstant(mod, "DICT_TYPE_AUTO", ZSTD_dct_auto);
+    PyModule_AddIntConstant(mod, "DICT_TYPE_RAWCONTENT", ZSTD_dct_rawContent);
+    PyModule_AddIntConstant(mod, "DICT_TYPE_FULLDICT", ZSTD_dct_fullDict);
+
+    PyModule_AddIntConstant(mod, "FORMAT_ZSTD1", ZSTD_f_zstd1);
+    PyModule_AddIntConstant(mod, "FORMAT_ZSTD1_MAGICLESS",
+                            ZSTD_f_zstd1_magicless);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/decompressionreader.c b/contrib/python/zstandard/py3/c-ext/decompressionreader.c
new file mode 100644
index 00000000000..0e0935356f1
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/decompressionreader.c
@@ -0,0 +1,780 @@
+/**
+ * Copyright (c) 2017-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void decompressionreader_dealloc(ZstdDecompressionReader *self) {
+    Py_XDECREF(self->decompressor);
+    Py_XDECREF(self->reader);
+
+    if (self->buffer.buf) {
+        PyBuffer_Release(&self->buffer);
+    }
+
+    Py_CLEAR(self->readResult);
+
+    PyObject_Del(self);
+}
+
+static ZstdDecompressionReader *
+decompressionreader_enter(ZstdDecompressionReader *self) {
+    if (self->entered) {
+        PyErr_SetString(PyExc_ValueError, "cannot __enter__ multiple times");
+        return NULL;
+    }
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    self->entered = 1;
+
+    Py_INCREF(self);
+    return self;
+}
+
+static PyObject *decompressionreader_exit(ZstdDecompressionReader *self,
+                                          PyObject *args) {
+    PyObject *exc_type;
+    PyObject *exc_value;
+    PyObject *exc_tb;
+
+    if (!PyArg_ParseTuple(args, "OOO:__exit__", &exc_type, &exc_value,
+                          &exc_tb)) {
+        return NULL;
+    }
+
+    self->entered = 0;
+
+    if (NULL == PyObject_CallMethod((PyObject *)self, "close", NULL)) {
+        return NULL;
+    }
+
+    /* Release resources. */
+    Py_CLEAR(self->reader);
+    if (self->buffer.buf) {
+        PyBuffer_Release(&self->buffer);
+        memset(&self->buffer, 0, sizeof(self->buffer));
+    }
+
+    Py_CLEAR(self->decompressor);
+
+    Py_RETURN_FALSE;
+}
+
+static PyObject *decompressionreader_readable(PyObject *self) {
+    Py_RETURN_TRUE;
+}
+
+static PyObject *decompressionreader_writable(PyObject *self) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *decompressionreader_seekable(PyObject *self) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *decompressionreader_close(ZstdDecompressionReader *self) {
+    if (self->closed) {
+        Py_RETURN_NONE;
+    }
+
+    self->closed = 1;
+
+    if (self->closefd && self->reader != NULL &&
+        PyObject_HasAttrString(self->reader, "close")) {
+        return PyObject_CallMethod(self->reader, "close", NULL);
+    }
+
+    Py_RETURN_NONE;
+}
+
+static PyObject *decompressionreader_flush(PyObject *self) {
+    Py_RETURN_NONE;
+}
+
+static PyObject *decompressionreader_isatty(PyObject *self) {
+    Py_RETURN_FALSE;
+}
+
+/**
+ * Read available input.
+ *
+ * Returns 0 if no data was added to input.
+ * Returns 1 if new input data is available.
+ * Returns -1 on error and sets a Python exception as a side-effect.
+ */
+int read_decompressor_input(ZstdDecompressionReader *self) {
+    if (self->finishedInput) {
+        return 0;
+    }
+
+    if (self->input.pos != self->input.size) {
+        return 0;
+    }
+
+    if (self->reader) {
+        Py_buffer buffer;
+
+        assert(self->readResult == NULL);
+        self->readResult =
+            PyObject_CallMethod(self->reader, "read", "k", self->readSize);
+        if (NULL == self->readResult) {
+            return -1;
+        }
+
+        memset(&buffer, 0, sizeof(buffer));
+
+        if (0 !=
+            PyObject_GetBuffer(self->readResult, &buffer, PyBUF_CONTIG_RO)) {
+            return -1;
+        }
+
+        /* EOF */
+        if (0 == buffer.len) {
+            self->finishedInput = 1;
+            Py_CLEAR(self->readResult);
+        }
+        else {
+            self->input.src = buffer.buf;
+            self->input.size = buffer.len;
+            self->input.pos = 0;
+        }
+
+        PyBuffer_Release(&buffer);
+    }
+    else {
+        assert(self->buffer.buf);
+        /*
+         * We should only get here once since expectation is we always
+         * exhaust input buffer before reading again.
+         */
+        assert(self->input.src == NULL);
+
+        self->input.src = self->buffer.buf;
+        self->input.size = self->buffer.len;
+        self->input.pos = 0;
+    }
+
+    return 1;
+}
+
+/**
+ * Decompresses available input into an output buffer.
+ *
+ * Returns 0 if we need more input.
+ * Returns 1 if output buffer should be emitted.
+ * Returns -1 on error and sets a Python exception.
+ */
+int decompress_input(ZstdDecompressionReader *self, ZSTD_outBuffer *output) {
+    size_t zresult;
+
+    if (self->input.pos >= self->input.size) {
+        return 0;
+    }
+
+    Py_BEGIN_ALLOW_THREADS zresult =
+        ZSTD_decompressStream(self->decompressor->dctx, output, &self->input);
+    Py_END_ALLOW_THREADS
+
+        /* Input exhausted. Clear our state tracking. */
+        if (self->input.pos == self->input.size) {
+        memset(&self->input, 0, sizeof(self->input));
+        Py_CLEAR(self->readResult);
+
+        if (self->buffer.buf) {
+            self->finishedInput = 1;
+        }
+    }
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "zstd decompress error: %s",
+                     ZSTD_getErrorName(zresult));
+        return -1;
+    }
+
+    /* We fulfilled the full read request. Signal to emit. */
+    if (output->pos && output->pos == output->size) {
+        return 1;
+    }
+    /* We're at the end of a frame and we aren't allowed to return data
+       spanning frames. */
+    else if (output->pos && zresult == 0 && !self->readAcrossFrames) {
+        return 1;
+    }
+
+    /* There is more room in the output. Signal to collect more data. */
+    return 0;
+}
+
+static PyObject *decompressionreader_read(ZstdDecompressionReader *self,
+                                          PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"size", NULL};
+
+    Py_ssize_t size = -1;
+    PyObject *result = NULL;
+    char *resultBuffer;
+    Py_ssize_t resultSize;
+    ZSTD_outBuffer output;
+    int decompressResult, readResult;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n", kwlist, &size)) {
+        return NULL;
+    }
+
+    if (size < -1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot read negative amounts less than -1");
+        return NULL;
+    }
+
+    if (size == -1) {
+        return PyObject_CallMethod((PyObject *)self, "readall", NULL);
+    }
+
+    if (self->finishedOutput || size == 0) {
+        return PyBytes_FromStringAndSize("", 0);
+    }
+
+    result = PyBytes_FromStringAndSize(NULL, size);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+    output.dst = resultBuffer;
+    output.size = resultSize;
+    output.pos = 0;
+
+readinput:
+
+    decompressResult = decompress_input(self, &output);
+
+    if (-1 == decompressResult) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+    else if (0 == decompressResult) {
+    }
+    else if (1 == decompressResult) {
+        self->bytesDecompressed += output.pos;
+
+        if (output.pos != output.size) {
+            if (safe_pybytes_resize(&result, output.pos)) {
+                Py_XDECREF(result);
+                return NULL;
+            }
+        }
+        return result;
+    }
+    else {
+        assert(0);
+    }
+
+    readResult = read_decompressor_input(self);
+
+    if (-1 == readResult) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+    else if (0 == readResult) {
+    }
+    else if (1 == readResult) {
+    }
+    else {
+        assert(0);
+    }
+
+    if (self->input.size) {
+        goto readinput;
+    }
+
+    /* EOF */
+    self->bytesDecompressed += output.pos;
+
+    if (safe_pybytes_resize(&result, output.pos)) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+
+    return result;
+}
+
+static PyObject *decompressionreader_read1(ZstdDecompressionReader *self,
+                                           PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"size", NULL};
+
+    Py_ssize_t size = -1;
+    PyObject *result = NULL;
+    char *resultBuffer;
+    Py_ssize_t resultSize;
+    ZSTD_outBuffer output;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|n", kwlist, &size)) {
+        return NULL;
+    }
+
+    if (size < -1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "cannot read negative amounts less than -1");
+        return NULL;
+    }
+
+    if (self->finishedOutput || size == 0) {
+        return PyBytes_FromStringAndSize("", 0);
+    }
+
+    if (size == -1) {
+        size = ZSTD_DStreamOutSize();
+    }
+
+    result = PyBytes_FromStringAndSize(NULL, size);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    PyBytes_AsStringAndSize(result, &resultBuffer, &resultSize);
+
+    output.dst = resultBuffer;
+    output.size = resultSize;
+    output.pos = 0;
+
+    /* read1() is supposed to use at most 1 read() from the underlying stream.
+     * However, we can't satisfy this requirement with decompression due to the
+     * nature of how decompression works. Our strategy is to read + decompress
+     * until we get any output, at which point we return. This satisfies the
+     * intent of the read1() API to limit read operations.
+     */
+    while (!self->finishedInput) {
+        int readResult, decompressResult;
+
+        readResult = read_decompressor_input(self);
+        if (-1 == readResult) {
+            Py_XDECREF(result);
+            return NULL;
+        }
+        else if (0 == readResult || 1 == readResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        decompressResult = decompress_input(self, &output);
+
+        if (-1 == decompressResult) {
+            Py_XDECREF(result);
+            return NULL;
+        }
+        else if (0 == decompressResult || 1 == decompressResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        if (output.pos) {
+            break;
+        }
+    }
+
+    self->bytesDecompressed += output.pos;
+    if (safe_pybytes_resize(&result, output.pos)) {
+        Py_XDECREF(result);
+        return NULL;
+    }
+
+    return result;
+}
+
+static PyObject *decompressionreader_readinto(ZstdDecompressionReader *self,
+                                              PyObject *args) {
+    Py_buffer dest;
+    ZSTD_outBuffer output;
+    int decompressResult, readResult;
+    PyObject *result = NULL;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (self->finishedOutput) {
+        return PyLong_FromLong(0);
+    }
+
+    if (!PyArg_ParseTuple(args, "w*:readinto", &dest)) {
+        return NULL;
+    }
+
+    output.dst = dest.buf;
+    output.size = dest.len;
+    output.pos = 0;
+
+readinput:
+
+    decompressResult = decompress_input(self, &output);
+
+    if (-1 == decompressResult) {
+        goto finally;
+    }
+    else if (0 == decompressResult) {
+    }
+    else if (1 == decompressResult) {
+        self->bytesDecompressed += output.pos;
+        result = PyLong_FromSize_t(output.pos);
+        goto finally;
+    }
+    else {
+        assert(0);
+    }
+
+    readResult = read_decompressor_input(self);
+
+    if (-1 == readResult) {
+        goto finally;
+    }
+    else if (0 == readResult) {
+    }
+    else if (1 == readResult) {
+    }
+    else {
+        assert(0);
+    }
+
+    if (self->input.size) {
+        goto readinput;
+    }
+
+    /* EOF */
+    self->bytesDecompressed += output.pos;
+    result = PyLong_FromSize_t(output.pos);
+
+finally:
+    PyBuffer_Release(&dest);
+
+    return result;
+}
+
+static PyObject *decompressionreader_readinto1(ZstdDecompressionReader *self,
+                                               PyObject *args) {
+    Py_buffer dest;
+    ZSTD_outBuffer output;
+    PyObject *result = NULL;
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (self->finishedOutput) {
+        return PyLong_FromLong(0);
+    }
+
+    if (!PyArg_ParseTuple(args, "w*:readinto1", &dest)) {
+        return NULL;
+    }
+
+    output.dst = dest.buf;
+    output.size = dest.len;
+    output.pos = 0;
+
+    while (!self->finishedInput && !self->finishedOutput) {
+        int decompressResult, readResult;
+
+        readResult = read_decompressor_input(self);
+
+        if (-1 == readResult) {
+            goto finally;
+        }
+        else if (0 == readResult || 1 == readResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        decompressResult = decompress_input(self, &output);
+
+        if (-1 == decompressResult) {
+            goto finally;
+        }
+        else if (0 == decompressResult || 1 == decompressResult) {
+        }
+        else {
+            assert(0);
+        }
+
+        if (output.pos) {
+            break;
+        }
+    }
+
+    self->bytesDecompressed += output.pos;
+    result = PyLong_FromSize_t(output.pos);
+
+finally:
+    PyBuffer_Release(&dest);
+
+    return result;
+}
+
+static PyObject *decompressionreader_readall(PyObject *self) {
+    PyObject *chunks = NULL;
+    PyObject *empty = NULL;
+    PyObject *result = NULL;
+
+    /* Our strategy is to collect chunks into a list then join all the
+     * chunks at the end. We could potentially use e.g. an io.BytesIO. But
+     * this feels simple enough to implement and avoids potentially expensive
+     * reallocations of large buffers.
+     */
+    chunks = PyList_New(0);
+    if (NULL == chunks) {
+        return NULL;
+    }
+
+    while (1) {
+        PyObject *chunk = PyObject_CallMethod(self, "read", "i", 1048576);
+        if (NULL == chunk) {
+            Py_DECREF(chunks);
+            return NULL;
+        }
+
+        if (!PyBytes_Size(chunk)) {
+            Py_DECREF(chunk);
+            break;
+        }
+
+        if (PyList_Append(chunks, chunk)) {
+            Py_DECREF(chunk);
+            Py_DECREF(chunks);
+            return NULL;
+        }
+
+        Py_DECREF(chunk);
+    }
+
+    empty = PyBytes_FromStringAndSize("", 0);
+    if (NULL == empty) {
+        Py_DECREF(chunks);
+        return NULL;
+    }
+
+    result = PyObject_CallMethod(empty, "join", "O", chunks);
+
+    Py_DECREF(empty);
+    Py_DECREF(chunks);
+
+    return result;
+}
+
+static PyObject *decompressionreader_readline(PyObject *self, PyObject *args,
+                                              PyObject *kwargs) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *decompressionreader_readlines(PyObject *self, PyObject *args,
+                                               PyObject *kwargs) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *decompressionreader_seek(ZstdDecompressionReader *self,
+                                          PyObject *args) {
+    Py_ssize_t pos;
+    int whence = 0;
+    unsigned long long readAmount = 0;
+    size_t defaultOutSize = ZSTD_DStreamOutSize();
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTuple(args, "n|i:seek", &pos, &whence)) {
+        return NULL;
+    }
+
+    if (whence == SEEK_SET) {
+        if (pos < 0) {
+            PyErr_SetString(PyExc_OSError,
+                            "cannot seek to negative position with SEEK_SET");
+            return NULL;
+        }
+
+        if ((unsigned long long)pos < self->bytesDecompressed) {
+            PyErr_SetString(PyExc_OSError,
+                            "cannot seek zstd decompression stream backwards");
+            return NULL;
+        }
+
+        readAmount = pos - self->bytesDecompressed;
+    }
+    else if (whence == SEEK_CUR) {
+        if (pos < 0) {
+            PyErr_SetString(PyExc_OSError,
+                            "cannot seek zstd decompression stream backwards");
+            return NULL;
+        }
+
+        readAmount = pos;
+    }
+    else if (whence == SEEK_END) {
+        /* We /could/ support this with pos==0. But let's not do that until
+           someone needs it. */
+        PyErr_SetString(
+            PyExc_OSError,
+            "zstd decompression streams cannot be seeked with SEEK_END");
+        return NULL;
+    }
+
+    /* It is a bit inefficient to do this via the Python API. But since there
+       is a bit of state tracking involved to read from this type, it is the
+       easiest to implement. */
+    while (readAmount) {
+        Py_ssize_t readSize;
+        PyObject *readResult = PyObject_CallMethod(
+            (PyObject *)self, "read", "K",
+            readAmount < defaultOutSize ? readAmount : defaultOutSize);
+
+        if (!readResult) {
+            return NULL;
+        }
+
+        readSize = PyBytes_GET_SIZE(readResult);
+
+        Py_CLEAR(readResult);
+
+        /* Empty read means EOF. */
+        if (!readSize) {
+            break;
+        }
+
+        readAmount -= readSize;
+    }
+
+    return PyLong_FromUnsignedLongLong(self->bytesDecompressed);
+}
+
+static PyObject *decompressionreader_tell(ZstdDecompressionReader *self) {
+    /* TODO should this raise OSError since stream isn't seekable? */
+    return PyLong_FromUnsignedLongLong(self->bytesDecompressed);
+}
+
+static PyObject *decompressionreader_write(PyObject *self, PyObject *args) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *decompressionreader_writelines(PyObject *self,
+                                                PyObject *args) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *decompressionreader_iter(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *decompressionreader_iternext(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyMethodDef decompressionreader_methods[] = {
+    {"__enter__", (PyCFunction)decompressionreader_enter, METH_NOARGS,
+     PyDoc_STR("Enter a compression context")},
+    {"__exit__", (PyCFunction)decompressionreader_exit, METH_VARARGS,
+     PyDoc_STR("Exit a compression context")},
+    {"close", (PyCFunction)decompressionreader_close, METH_NOARGS,
+     PyDoc_STR("Close the stream so it cannot perform any more operations")},
+    {"flush", (PyCFunction)decompressionreader_flush, METH_NOARGS,
+     PyDoc_STR("no-ops")},
+    {"isatty", (PyCFunction)decompressionreader_isatty, METH_NOARGS,
+     PyDoc_STR("Returns False")},
+    {"readable", (PyCFunction)decompressionreader_readable, METH_NOARGS,
+     PyDoc_STR("Returns True")},
+    {"read", (PyCFunction)decompressionreader_read,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("read compressed data")},
+    {"read1", (PyCFunction)decompressionreader_read1,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("read compressed data")},
+    {"readinto", (PyCFunction)decompressionreader_readinto, METH_VARARGS, NULL},
+    {"readinto1", (PyCFunction)decompressionreader_readinto1, METH_VARARGS,
+     NULL},
+    {"readall", (PyCFunction)decompressionreader_readall, METH_NOARGS,
+     PyDoc_STR("Not implemented")},
+    {"readline", (PyCFunction)decompressionreader_readline,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("Not implemented")},
+    {"readlines", (PyCFunction)decompressionreader_readlines,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("Not implemented")},
+    {"seek", (PyCFunction)decompressionreader_seek, METH_VARARGS,
+     PyDoc_STR("Seek the stream")},
+    {"seekable", (PyCFunction)decompressionreader_seekable, METH_NOARGS,
+     PyDoc_STR("Returns False")},
+    {"tell", (PyCFunction)decompressionreader_tell, METH_NOARGS,
+     PyDoc_STR("Returns current number of bytes compressed")},
+    {"writable", (PyCFunction)decompressionreader_writable, METH_NOARGS,
+     PyDoc_STR("Returns False")},
+    {"write", (PyCFunction)decompressionreader_write, METH_VARARGS,
+     PyDoc_STR("unsupported operation")},
+    {"writelines", (PyCFunction)decompressionreader_writelines, METH_VARARGS,
+     PyDoc_STR("unsupported operation")},
+    {NULL, NULL}};
+
+static PyMemberDef decompressionreader_members[] = {
+    {"closed", T_BOOL, offsetof(ZstdDecompressionReader, closed), READONLY,
+     "whether stream is closed"},
+    {NULL}};
+
+PyType_Slot ZstdDecompressionReaderSlots[] = {
+    {Py_tp_dealloc, decompressionreader_dealloc},
+    {Py_tp_iter, decompressionreader_iter},
+    {Py_tp_iternext, decompressionreader_iternext},
+    {Py_tp_methods, decompressionreader_methods},
+    {Py_tp_members, decompressionreader_members},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdDecompressionReaderSpec = {
+    "zstd.ZstdDecompressionReader",
+    sizeof(ZstdDecompressionReader),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    ZstdDecompressionReaderSlots,
+};
+
+PyTypeObject *ZstdDecompressionReaderType;
+
+void decompressionreader_module_init(PyObject *mod) {
+    /* TODO make reader a sub-class of io.RawIOBase */
+
+    ZstdDecompressionReaderType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdDecompressionReaderSpec);
+    if (PyType_Ready(ZstdDecompressionReaderType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdDecompressionReaderType);
+    PyModule_AddObject(mod, "ZstdDecompressionReader",
+                       (PyObject *)ZstdDecompressionReaderType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/decompressionwriter.c b/contrib/python/zstandard/py3/c-ext/decompressionwriter.c
new file mode 100644
index 00000000000..09fd1338056
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/decompressionwriter.c
@@ -0,0 +1,273 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void ZstdDecompressionWriter_dealloc(ZstdDecompressionWriter *self) {
+    Py_XDECREF(self->decompressor);
+    Py_XDECREF(self->writer);
+
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdDecompressionWriter_enter(ZstdDecompressionWriter *self) {
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (self->entered) {
+        PyErr_SetString(ZstdError, "cannot __enter__ multiple times");
+        return NULL;
+    }
+
+    self->entered = 1;
+
+    Py_INCREF(self);
+    return (PyObject *)self;
+}
+
+static PyObject *ZstdDecompressionWriter_exit(ZstdDecompressionWriter *self,
+                                              PyObject *args) {
+    self->entered = 0;
+
+    if (NULL == PyObject_CallMethod((PyObject *)self, "close", NULL)) {
+        return NULL;
+    }
+
+    Py_RETURN_FALSE;
+}
+
+static PyObject *
+ZstdDecompressionWriter_memory_size(ZstdDecompressionWriter *self) {
+    return PyLong_FromSize_t(ZSTD_sizeof_DCtx(self->decompressor->dctx));
+}
+
+static PyObject *ZstdDecompressionWriter_write(ZstdDecompressionWriter *self,
+                                               PyObject *args,
+                                               PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    PyObject *result = NULL;
+    Py_buffer source;
+    size_t zresult = 0;
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    PyObject *res;
+    Py_ssize_t totalWrite = 0;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:write", kwlist,
+                                     &source)) {
+        return NULL;
+    }
+
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    output.dst = PyMem_Malloc(self->outSize);
+    if (!output.dst) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+    output.size = self->outSize;
+    output.pos = 0;
+
+    input.src = source.buf;
+    input.size = source.len;
+    input.pos = 0;
+
+    while (input.pos < (size_t)source.len) {
+        Py_BEGIN_ALLOW_THREADS zresult =
+            ZSTD_decompressStream(self->decompressor->dctx, &output, &input);
+        Py_END_ALLOW_THREADS
+
+            if (ZSTD_isError(zresult)) {
+            PyMem_Free(output.dst);
+            PyErr_Format(ZstdError, "zstd decompress error: %s",
+                         ZSTD_getErrorName(zresult));
+            goto finally;
+        }
+
+        if (output.pos) {
+            res = PyObject_CallMethod(self->writer, "write", "y#", output.dst,
+                                      output.pos);
+            if (NULL == res) {
+                goto finally;
+            }
+            Py_XDECREF(res);
+            totalWrite += output.pos;
+            output.pos = 0;
+        }
+    }
+
+    PyMem_Free(output.dst);
+
+    if (self->writeReturnRead) {
+        result = PyLong_FromSize_t(input.pos);
+    }
+    else {
+        result = PyLong_FromSsize_t(totalWrite);
+    }
+
+finally:
+    PyBuffer_Release(&source);
+    return result;
+}
+
+static PyObject *ZstdDecompressionWriter_close(ZstdDecompressionWriter *self) {
+    PyObject *result;
+
+    if (self->closed) {
+        Py_RETURN_NONE;
+    }
+
+    self->closing = 1;
+    result = PyObject_CallMethod((PyObject *)self, "flush", NULL);
+    self->closing = 0;
+    self->closed = 1;
+
+    if (NULL == result) {
+        return NULL;
+    }
+
+    /* Call close on underlying stream as well. */
+    if (self->closefd && PyObject_HasAttrString(self->writer, "close")) {
+        return PyObject_CallMethod(self->writer, "close", NULL);
+    }
+
+    Py_RETURN_NONE;
+}
+
+static PyObject *ZstdDecompressionWriter_fileno(ZstdDecompressionWriter *self) {
+    if (PyObject_HasAttrString(self->writer, "fileno")) {
+        return PyObject_CallMethod(self->writer, "fileno", NULL);
+    }
+    else {
+        PyErr_SetString(PyExc_OSError,
+                        "fileno not available on underlying writer");
+        return NULL;
+    }
+}
+
+static PyObject *ZstdDecompressionWriter_flush(ZstdDecompressionWriter *self) {
+    if (self->closed) {
+        PyErr_SetString(PyExc_ValueError, "stream is closed");
+        return NULL;
+    }
+
+    if (!self->closing && PyObject_HasAttrString(self->writer, "flush")) {
+        return PyObject_CallMethod(self->writer, "flush", NULL);
+    }
+    else {
+        Py_RETURN_NONE;
+    }
+}
+
+static PyObject *ZstdDecompressionWriter_iter(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *ZstdDecompressionWriter_iternext(PyObject *self) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyObject *ZstdDecompressionWriter_false(PyObject *self, PyObject *args) {
+    Py_RETURN_FALSE;
+}
+
+static PyObject *ZstdDecompressionWriter_true(PyObject *self, PyObject *args) {
+    Py_RETURN_TRUE;
+}
+
+static PyObject *ZstdDecompressionWriter_unsupported(PyObject *self,
+                                                     PyObject *args,
+                                                     PyObject *kwargs) {
+    set_io_unsupported_operation();
+    return NULL;
+}
+
+static PyMethodDef ZstdDecompressionWriter_methods[] = {
+    {"__enter__", (PyCFunction)ZstdDecompressionWriter_enter, METH_NOARGS,
+     PyDoc_STR("Enter a decompression context.")},
+    {"__exit__", (PyCFunction)ZstdDecompressionWriter_exit, METH_VARARGS,
+     PyDoc_STR("Exit a decompression context.")},
+    {"memory_size", (PyCFunction)ZstdDecompressionWriter_memory_size,
+     METH_NOARGS,
+     PyDoc_STR(
+         "Obtain the memory size in bytes of the underlying decompressor.")},
+    {"close", (PyCFunction)ZstdDecompressionWriter_close, METH_NOARGS, NULL},
+    {"fileno", (PyCFunction)ZstdDecompressionWriter_fileno, METH_NOARGS, NULL},
+    {"flush", (PyCFunction)ZstdDecompressionWriter_flush, METH_NOARGS, NULL},
+    {"isatty", ZstdDecompressionWriter_false, METH_NOARGS, NULL},
+    {"readable", ZstdDecompressionWriter_false, METH_NOARGS, NULL},
+    {"readline", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readlines", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"seek", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"seekable", ZstdDecompressionWriter_false, METH_NOARGS, NULL},
+    {"tell", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"truncate", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"writable", ZstdDecompressionWriter_true, METH_NOARGS, NULL},
+    {"writelines", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"read", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readall", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"readinto", (PyCFunction)ZstdDecompressionWriter_unsupported,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"write", (PyCFunction)ZstdDecompressionWriter_write,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("Compress data")},
+    {NULL, NULL}};
+
+static PyMemberDef ZstdDecompressionWriter_members[] = {
+    {"closed", T_BOOL, offsetof(ZstdDecompressionWriter, closed), READONLY,
+     NULL},
+    {NULL}};
+
+PyType_Slot ZstdDecompressionWriterSlots[] = {
+    {Py_tp_dealloc, ZstdDecompressionWriter_dealloc},
+    {Py_tp_iter, ZstdDecompressionWriter_iter},
+    {Py_tp_iternext, ZstdDecompressionWriter_iternext},
+    {Py_tp_methods, ZstdDecompressionWriter_methods},
+    {Py_tp_members, ZstdDecompressionWriter_members},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdDecompressionWriterSpec = {
+    "zstd.ZstdDecompressionWriter",
+    sizeof(ZstdDecompressionWriter),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdDecompressionWriterSlots,
+};
+
+PyTypeObject *ZstdDecompressionWriterType;
+
+void decompressionwriter_module_init(PyObject *mod) {
+    ZstdDecompressionWriterType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdDecompressionWriterSpec);
+    if (PyType_Ready(ZstdDecompressionWriterType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdDecompressionWriterType);
+    PyModule_AddObject(mod, "ZstdDecompressionWriter",
+                       (PyObject *)ZstdDecompressionWriterType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/decompressobj.c b/contrib/python/zstandard/py3/c-ext/decompressobj.c
new file mode 100644
index 00000000000..5c3a124fc9a
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/decompressobj.c
@@ -0,0 +1,201 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+static void DecompressionObj_dealloc(ZstdDecompressionObj *self) {
+    Py_XDECREF(self->decompressor);
+    Py_CLEAR(self->unused_data);
+
+    PyObject_Del(self);
+}
+
+static PyObject *DecompressionObj_decompress(ZstdDecompressionObj *self,
+                                             PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    Py_buffer source;
+    size_t zresult;
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    PyObject *result = NULL;
+    Py_ssize_t resultSize = 0;
+
+    output.dst = NULL;
+
+    if (self->finished) {
+        PyErr_SetString(ZstdError, "cannot use a decompressobj multiple times");
+        return NULL;
+    }
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:decompress", kwlist,
+                                     &source)) {
+        return NULL;
+    }
+
+    /* Special case of empty input. Output will always be empty. */
+    if (source.len == 0) {
+        result = PyBytes_FromString("");
+        goto finally;
+    }
+
+    input.src = source.buf;
+    input.size = source.len;
+    input.pos = 0;
+
+    output.dst = PyMem_Malloc(self->outSize);
+    if (!output.dst) {
+        PyErr_NoMemory();
+        goto except;
+    }
+    output.size = self->outSize;
+    output.pos = 0;
+
+    while (1) {
+        Py_BEGIN_ALLOW_THREADS zresult =
+            ZSTD_decompressStream(self->decompressor->dctx, &output, &input);
+        Py_END_ALLOW_THREADS
+
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "zstd decompressor error: %s",
+                         ZSTD_getErrorName(zresult));
+            goto except;
+        }
+
+        if (output.pos) {
+            if (result) {
+                resultSize = PyBytes_GET_SIZE(result);
+                if (-1 ==
+                    safe_pybytes_resize(&result, resultSize + output.pos)) {
+                    Py_XDECREF(result);
+                    goto except;
+                }
+
+                memcpy(PyBytes_AS_STRING(result) + resultSize, output.dst,
+                       output.pos);
+            }
+            else {
+                result = PyBytes_FromStringAndSize(output.dst, output.pos);
+                if (!result) {
+                    goto except;
+                }
+            }
+        }
+
+        if (0 == zresult) {
+            self->finished = 1;
+
+            /* We should only get here at most once. */
+            assert(!self->unused_data);
+            self->unused_data = PyBytes_FromStringAndSize((char *)(input.src) + input.pos, input.size - input.pos);
+
+            break;
+        }
+        else if (input.pos == input.size && output.pos == 0) {
+            break;
+        }
+        else {
+            output.pos = 0;
+        }
+    }
+
+    if (!result) {
+        result = PyBytes_FromString("");
+    }
+
+    goto finally;
+
+except:
+    Py_CLEAR(result);
+
+finally:
+    PyMem_Free(output.dst);
+    PyBuffer_Release(&source);
+
+    return result;
+}
+
+static PyObject *DecompressionObj_flush(ZstdDecompressionObj *self,
+                                        PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"length", NULL};
+
+    PyObject *length = NULL;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|O:flush", kwlist,
+                                     &length)) {
+        return NULL;
+    }
+
+    return PyBytes_FromString("");
+}
+
+static PyObject *DecompressionObj_unused_data(PyObject *self, void *unused) {
+    ZstdDecompressionObj *slf = (ZstdDecompressionObj *)(self);
+
+    if (slf->unused_data) {
+        Py_INCREF(slf->unused_data);
+        return slf->unused_data;
+    } else {
+        return PyBytes_FromString("");
+    }
+}
+
+static PyObject *DecompressionObj_unconsumed_tail(PyObject *self, void *unused) {
+    return PyBytes_FromString("");
+}
+
+static PyObject *DecompressionObj_eof(PyObject *self, void *unused) {
+    ZstdDecompressionObj *slf = (ZstdDecompressionObj*)(self);
+
+    if (slf->finished) {
+        Py_RETURN_TRUE;
+    } else {
+        Py_RETURN_FALSE;
+    }
+}
+
+static PyMethodDef DecompressionObj_methods[] = {
+    {"decompress", (PyCFunction)DecompressionObj_decompress,
+     METH_VARARGS | METH_KEYWORDS, PyDoc_STR("decompress data")},
+    {"flush", (PyCFunction)DecompressionObj_flush, METH_VARARGS | METH_KEYWORDS,
+     PyDoc_STR("no-op")},
+    {NULL, NULL}};
+
+static PyGetSetDef DecompressionObj_getset[] = {
+    {"unused_data", DecompressionObj_unused_data, NULL, NULL, NULL },
+    {"unconsumed_tail", DecompressionObj_unconsumed_tail, NULL, NULL, NULL },
+    {"eof", DecompressionObj_eof, NULL, NULL, NULL },
+    {NULL}};
+
+PyType_Slot ZstdDecompressionObjSlots[] = {
+    {Py_tp_dealloc, DecompressionObj_dealloc},
+    {Py_tp_methods, DecompressionObj_methods},
+    {Py_tp_getset, DecompressionObj_getset},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdDecompressionObjSpec = {
+    "zstd.ZstdDecompressionObj",
+    sizeof(ZstdDecompressionObj),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdDecompressionObjSlots,
+};
+
+PyTypeObject *ZstdDecompressionObjType;
+
+void decompressobj_module_init(PyObject *module) {
+    ZstdDecompressionObjType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdDecompressionObjSpec);
+    if (PyType_Ready(ZstdDecompressionObjType) < 0) {
+        return;
+    }
+}
diff --git a/contrib/python/zstandard/py3/c-ext/decompressor.c b/contrib/python/zstandard/py3/c-ext/decompressor.c
new file mode 100644
index 00000000000..3e61fe9ee6f
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/decompressor.c
@@ -0,0 +1,1762 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+/**
+ * Ensure the ZSTD_DCtx on a decompressor is initiated and ready for a new
+ * operation.
+ */
+int ensure_dctx(ZstdDecompressor *decompressor, int loadDict) {
+    size_t zresult;
+
+    ZSTD_DCtx_reset(decompressor->dctx, ZSTD_reset_session_only);
+
+    if (decompressor->maxWindowSize) {
+        zresult = ZSTD_DCtx_setMaxWindowSize(decompressor->dctx,
+                                             decompressor->maxWindowSize);
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError, "unable to set max window size: %s",
+                         ZSTD_getErrorName(zresult));
+            return 1;
+        }
+    }
+
+    zresult = ZSTD_DCtx_setParameter(decompressor->dctx, ZSTD_d_format,
+                                     decompressor->format);
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "unable to set decoding format: %s",
+                     ZSTD_getErrorName(zresult));
+        return 1;
+    }
+
+    if (loadDict && decompressor->dict) {
+        if (ensure_ddict(decompressor->dict)) {
+            return 1;
+        }
+
+        zresult =
+            ZSTD_DCtx_refDDict(decompressor->dctx, decompressor->dict->ddict);
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(ZstdError,
+                         "unable to reference prepared dictionary: %s",
+                         ZSTD_getErrorName(zresult));
+            return 1;
+        }
+    }
+
+    return 0;
+}
+
+static int Decompressor_init(ZstdDecompressor *self, PyObject *args,
+                             PyObject *kwargs) {
+    static char *kwlist[] = {"dict_data", "max_window_size", "format", NULL};
+
+    PyObject *dict = NULL;
+    Py_ssize_t maxWindowSize = 0;
+    ZSTD_format_e format = ZSTD_f_zstd1;
+
+    self->dctx = NULL;
+    self->dict = NULL;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|OnI:ZstdDecompressor",
+                                     kwlist, &dict, &maxWindowSize, &format)) {
+        return -1;
+    }
+
+    if (dict) {
+        if (dict == Py_None) {
+            dict = NULL;
+        }
+        else if (!PyObject_IsInstance(dict,
+                                      (PyObject *)ZstdCompressionDictType)) {
+            PyErr_Format(PyExc_TypeError,
+                         "dict_data must be zstd.ZstdCompressionDict");
+            return -1;
+        }
+    }
+
+    self->dctx = ZSTD_createDCtx();
+    if (!self->dctx) {
+        PyErr_NoMemory();
+        goto except;
+    }
+
+    self->maxWindowSize = maxWindowSize;
+    self->format = format;
+
+    if (dict) {
+        self->dict = (ZstdCompressionDict *)dict;
+        Py_INCREF(dict);
+    }
+
+    if (ensure_dctx(self, 1)) {
+        goto except;
+    }
+
+    return 0;
+
+except:
+    Py_CLEAR(self->dict);
+
+    if (self->dctx) {
+        ZSTD_freeDCtx(self->dctx);
+        self->dctx = NULL;
+    }
+
+    return -1;
+}
+
+static void Decompressor_dealloc(ZstdDecompressor *self) {
+    Py_CLEAR(self->dict);
+
+    if (self->dctx) {
+        ZSTD_freeDCtx(self->dctx);
+        self->dctx = NULL;
+    }
+
+    PyObject_Del(self);
+}
+
+static PyObject *Decompressor_memory_size(ZstdDecompressor *self) {
+    if (self->dctx) {
+        return PyLong_FromSize_t(ZSTD_sizeof_DCtx(self->dctx));
+    }
+    else {
+        PyErr_SetString(
+            ZstdError,
+            "no decompressor context found; this should never happen");
+        return NULL;
+    }
+}
+
+static PyObject *Decompressor_copy_stream(ZstdDecompressor *self,
+                                          PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"ifh", "ofh", "read_size", "write_size", NULL};
+
+    PyObject *source;
+    PyObject *dest;
+    size_t inSize = ZSTD_DStreamInSize();
+    size_t outSize = ZSTD_DStreamOutSize();
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    Py_ssize_t totalRead = 0;
+    Py_ssize_t totalWrite = 0;
+    char *readBuffer;
+    Py_ssize_t readSize;
+    PyObject *readResult = NULL;
+    PyObject *res = NULL;
+    size_t zresult = 0;
+    PyObject *writeResult;
+    PyObject *totalReadPy;
+    PyObject *totalWritePy;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "OO|kk:copy_stream", kwlist,
+                                     &source, &dest, &inSize, &outSize)) {
+        return NULL;
+    }
+
+    if (!PyObject_HasAttrString(source, "read")) {
+        PyErr_SetString(PyExc_ValueError,
+                        "first argument must have a read() method");
+        return NULL;
+    }
+
+    if (!PyObject_HasAttrString(dest, "write")) {
+        PyErr_SetString(PyExc_ValueError,
+                        "second argument must have a write() method");
+        return NULL;
+    }
+
+    /* Prevent free on uninitialized memory in finally. */
+    output.dst = NULL;
+
+    if (ensure_dctx(self, 1)) {
+        res = NULL;
+        goto finally;
+    }
+
+    output.dst = PyMem_Malloc(outSize);
+    if (!output.dst) {
+        PyErr_NoMemory();
+        res = NULL;
+        goto finally;
+    }
+    output.size = outSize;
+    output.pos = 0;
+
+    /* Read source stream until EOF */
+    while (1) {
+        readResult = PyObject_CallMethod(source, "read", "n", inSize);
+        if (!readResult) {
+            goto finally;
+        }
+
+        PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+
+        /* If no data was read, we're at EOF. */
+        if (0 == readSize) {
+            break;
+        }
+
+        totalRead += readSize;
+
+        /* Send data to decompressor */
+        input.src = readBuffer;
+        input.size = readSize;
+        input.pos = 0;
+
+        while (input.pos < input.size) {
+            Py_BEGIN_ALLOW_THREADS zresult =
+                ZSTD_decompressStream(self->dctx, &output, &input);
+            Py_END_ALLOW_THREADS
+
+                if (ZSTD_isError(zresult)) {
+                PyErr_Format(ZstdError, "zstd decompressor error: %s",
+                             ZSTD_getErrorName(zresult));
+                res = NULL;
+                goto finally;
+            }
+
+            if (output.pos) {
+                writeResult = PyObject_CallMethod(dest, "write", "y#",
+                                                  output.dst, output.pos);
+                if (NULL == writeResult) {
+                    res = NULL;
+                    goto finally;
+                }
+
+                Py_XDECREF(writeResult);
+                totalWrite += output.pos;
+                output.pos = 0;
+            }
+        }
+
+        Py_CLEAR(readResult);
+    }
+
+    /* Source stream is exhausted. Finish up. */
+
+    totalReadPy = PyLong_FromSsize_t(totalRead);
+    totalWritePy = PyLong_FromSsize_t(totalWrite);
+    res = PyTuple_Pack(2, totalReadPy, totalWritePy);
+    Py_DECREF(totalReadPy);
+    Py_DECREF(totalWritePy);
+
+finally:
+    if (output.dst) {
+        PyMem_Free(output.dst);
+    }
+
+    Py_XDECREF(readResult);
+
+    return res;
+}
+
+PyObject *Decompressor_decompress(ZstdDecompressor *self, PyObject *args,
+                                  PyObject *kwargs) {
+    static char *kwlist[] = {
+        "data",
+        "max_output_size",
+        "read_across_frames",
+        "allow_extra_data",
+        NULL
+    };
+
+    Py_buffer source;
+    Py_ssize_t maxOutputSize = 0;
+
+    unsigned long long decompressedSize;
+    PyObject *readAcrossFrames = NULL;
+    PyObject *allowExtraData = NULL;
+    size_t destCapacity;
+    PyObject *result = NULL;
+    size_t zresult;
+    ZSTD_outBuffer outBuffer;
+    ZSTD_inBuffer inBuffer;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|nOO:decompress", kwlist,
+                                     &source, &maxOutputSize, &readAcrossFrames,
+                                     &allowExtraData)) {
+        return NULL;
+    }
+
+    if (readAcrossFrames ? PyObject_IsTrue(readAcrossFrames) : 0) {
+        PyErr_SetString(ZstdError,
+            "ZstdDecompressor.read_across_frames=True is not yet implemented"
+        );
+        goto finally;
+    }
+
+    if (ensure_dctx(self, 1)) {
+        goto finally;
+    }
+
+    decompressedSize = ZSTD_getFrameContentSize(source.buf, source.len);
+
+    if (ZSTD_CONTENTSIZE_ERROR == decompressedSize) {
+        PyErr_SetString(ZstdError,
+                        "error determining content size from frame header");
+        goto finally;
+    }
+    /* Special case of empty frame. */
+    else if (0 == decompressedSize) {
+        result = PyBytes_FromStringAndSize("", 0);
+        goto finally;
+    }
+    /* Missing content size in frame header. */
+    if (ZSTD_CONTENTSIZE_UNKNOWN == decompressedSize) {
+        if (0 == maxOutputSize) {
+            PyErr_SetString(ZstdError,
+                            "could not determine content size in frame header");
+            goto finally;
+        }
+
+        result = PyBytes_FromStringAndSize(NULL, maxOutputSize);
+        destCapacity = maxOutputSize;
+        decompressedSize = 0;
+    }
+    /* Size is recorded in frame header. */
+    else {
+        assert(SIZE_MAX >= PY_SSIZE_T_MAX);
+        if (decompressedSize > PY_SSIZE_T_MAX) {
+            PyErr_SetString(
+                ZstdError, "frame is too large to decompress on this platform");
+            goto finally;
+        }
+
+        result = PyBytes_FromStringAndSize(NULL, (Py_ssize_t)decompressedSize);
+        destCapacity = (size_t)decompressedSize;
+    }
+
+    if (!result) {
+        goto finally;
+    }
+
+    outBuffer.dst = PyBytes_AsString(result);
+    outBuffer.size = destCapacity;
+    outBuffer.pos = 0;
+
+    inBuffer.src = source.buf;
+    inBuffer.size = source.len;
+    inBuffer.pos = 0;
+
+    Py_BEGIN_ALLOW_THREADS zresult =
+        ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+    Py_END_ALLOW_THREADS
+
+        if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "decompression error: %s",
+                     ZSTD_getErrorName(zresult));
+        Py_CLEAR(result);
+        goto finally;
+    }
+    else if (zresult) {
+        PyErr_Format(ZstdError,
+                     "decompression error: did not decompress full frame");
+        Py_CLEAR(result);
+        goto finally;
+    }
+    else if (decompressedSize && outBuffer.pos != decompressedSize) {
+        PyErr_Format(
+            ZstdError,
+            "decompression error: decompressed %zu bytes; expected %llu",
+            zresult, decompressedSize);
+        Py_CLEAR(result);
+        goto finally;
+    }
+    else if (outBuffer.pos < destCapacity) {
+        if (safe_pybytes_resize(&result, outBuffer.pos)) {
+            Py_CLEAR(result);
+            goto finally;
+        }
+    }
+    else if ((allowExtraData ? PyObject_IsTrue(allowExtraData) : 1) == 0
+             && inBuffer.pos < inBuffer.size) {
+        PyErr_Format(
+            ZstdError,
+            "compressed input contains %zu bytes of unused data, which is disallowed",
+            inBuffer.size - inBuffer.pos
+        );
+        Py_CLEAR(result);
+        goto finally;
+    }
+
+finally:
+    PyBuffer_Release(&source);
+    return result;
+}
+
+static ZstdDecompressionObj *Decompressor_decompressobj(ZstdDecompressor *self,
+                                                        PyObject *args,
+                                                        PyObject *kwargs) {
+    static char *kwlist[] = {"write_size", NULL};
+
+    ZstdDecompressionObj *result = NULL;
+    size_t outSize = ZSTD_DStreamOutSize();
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|k:decompressobj", kwlist,
+                                     &outSize)) {
+        return NULL;
+    }
+
+    if (!outSize) {
+        PyErr_SetString(PyExc_ValueError, "write_size must be positive");
+        return NULL;
+    }
+
+    result = (ZstdDecompressionObj *)PyObject_CallObject(
+        (PyObject *)ZstdDecompressionObjType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    if (ensure_dctx(self, 1)) {
+        Py_DECREF(result);
+        return NULL;
+    }
+
+    result->decompressor = self;
+    Py_INCREF(result->decompressor);
+    result->outSize = outSize;
+
+    return result;
+}
+
+static ZstdDecompressorIterator *
+Decompressor_read_to_iter(ZstdDecompressor *self, PyObject *args,
+                          PyObject *kwargs) {
+    static char *kwlist[] = {"reader", "read_size", "write_size", "skip_bytes",
+                             NULL};
+
+    PyObject *reader;
+    size_t inSize = ZSTD_DStreamInSize();
+    size_t outSize = ZSTD_DStreamOutSize();
+    ZstdDecompressorIterator *result;
+    size_t skipBytes = 0;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kkk:read_to_iter", kwlist,
+                                     &reader, &inSize, &outSize, &skipBytes)) {
+        return NULL;
+    }
+
+    if (skipBytes >= inSize) {
+        PyErr_SetString(PyExc_ValueError,
+                        "skip_bytes must be smaller than read_size");
+        return NULL;
+    }
+
+    result = (ZstdDecompressorIterator *)PyObject_CallObject(
+        (PyObject *)ZstdDecompressorIteratorType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    if (PyObject_HasAttrString(reader, "read")) {
+        result->reader = reader;
+        Py_INCREF(result->reader);
+    }
+    else if (1 == PyObject_CheckBuffer(reader)) {
+        /* Object claims it is a buffer. Try to get a handle to it. */
+        if (0 != PyObject_GetBuffer(reader, &result->buffer, PyBUF_CONTIG_RO)) {
+            goto except;
+        }
+    }
+    else {
+        PyErr_SetString(PyExc_ValueError,
+                        "must pass an object with a read() method or conforms "
+                        "to buffer protocol");
+        goto except;
+    }
+
+    result->decompressor = self;
+    Py_INCREF(result->decompressor);
+
+    result->inSize = inSize;
+    result->outSize = outSize;
+    result->skipBytes = skipBytes;
+
+    if (ensure_dctx(self, 1)) {
+        goto except;
+    }
+
+    result->input.src = PyMem_Malloc(inSize);
+    if (!result->input.src) {
+        PyErr_NoMemory();
+        goto except;
+    }
+
+    goto finally;
+
+except:
+    Py_CLEAR(result);
+
+finally:
+
+    return result;
+}
+
+static ZstdDecompressionReader *
+Decompressor_stream_reader(ZstdDecompressor *self, PyObject *args,
+                           PyObject *kwargs) {
+    static char *kwlist[] = {"source", "read_size", "read_across_frames",
+                             "closefd", NULL};
+
+    PyObject *source;
+    size_t readSize = ZSTD_DStreamInSize();
+    PyObject *readAcrossFrames = NULL;
+    PyObject *closefd = NULL;
+    ZstdDecompressionReader *result;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kOO:stream_reader",
+                                     kwlist, &source, &readSize,
+                                     &readAcrossFrames, &closefd)) {
+        return NULL;
+    }
+
+    if (ensure_dctx(self, 1)) {
+        return NULL;
+    }
+
+    result = (ZstdDecompressionReader *)PyObject_CallObject(
+        (PyObject *)ZstdDecompressionReaderType, NULL);
+    if (NULL == result) {
+        return NULL;
+    }
+
+    result->entered = 0;
+    result->closed = 0;
+
+    if (PyObject_HasAttrString(source, "read")) {
+        result->reader = source;
+        Py_INCREF(source);
+        result->readSize = readSize;
+    }
+    else if (1 == PyObject_CheckBuffer(source)) {
+        if (0 != PyObject_GetBuffer(source, &result->buffer, PyBUF_CONTIG_RO)) {
+            Py_CLEAR(result);
+            return NULL;
+        }
+    }
+    else {
+        PyErr_SetString(PyExc_TypeError,
+                        "must pass an object with a read() method or that "
+                        "conforms to the buffer protocol");
+        Py_CLEAR(result);
+        return NULL;
+    }
+
+    result->decompressor = self;
+    Py_INCREF(self);
+    result->readAcrossFrames =
+        readAcrossFrames ? PyObject_IsTrue(readAcrossFrames) : 0;
+    result->closefd = closefd ? PyObject_IsTrue(closefd) : 1;
+
+    return result;
+}
+
+static ZstdDecompressionWriter *
+Decompressor_stream_writer(ZstdDecompressor *self, PyObject *args,
+                           PyObject *kwargs) {
+    static char *kwlist[] = {"writer", "write_size", "write_return_read",
+                             "closefd", NULL};
+
+    PyObject *writer;
+    size_t outSize = ZSTD_DStreamOutSize();
+    PyObject *writeReturnRead = NULL;
+    PyObject *closefd = NULL;
+    ZstdDecompressionWriter *result;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kOO:stream_writer",
+                                     kwlist, &writer, &outSize,
+                                     &writeReturnRead, &closefd)) {
+        return NULL;
+    }
+
+    if (!PyObject_HasAttrString(writer, "write")) {
+        PyErr_SetString(PyExc_ValueError,
+                        "must pass an object with a write() method");
+        return NULL;
+    }
+
+    if (ensure_dctx(self, 1)) {
+        return NULL;
+    }
+
+    result = (ZstdDecompressionWriter *)PyObject_CallObject(
+        (PyObject *)ZstdDecompressionWriterType, NULL);
+    if (!result) {
+        return NULL;
+    }
+
+    result->entered = 0;
+    result->closing = 0;
+    result->closed = 0;
+
+    result->decompressor = self;
+    Py_INCREF(result->decompressor);
+
+    result->writer = writer;
+    Py_INCREF(result->writer);
+
+    result->outSize = outSize;
+    result->writeReturnRead =
+        writeReturnRead ? PyObject_IsTrue(writeReturnRead) : 1;
+    result->closefd = closefd ? PyObject_IsTrue(closefd) : 1;
+
+    return result;
+}
+
+static PyObject *
+Decompressor_decompress_content_dict_chain(ZstdDecompressor *self,
+                                           PyObject *args, PyObject *kwargs) {
+    static char *kwlist[] = {"frames", NULL};
+
+    PyObject *chunks;
+    Py_ssize_t chunksLen;
+    Py_ssize_t chunkIndex;
+    char parity = 0;
+    PyObject *chunk;
+    char *chunkData;
+    Py_ssize_t chunkSize;
+    size_t zresult;
+    ZSTD_frameHeader frameHeader;
+    void *buffer1 = NULL;
+    size_t buffer1Size = 0;
+    size_t buffer1ContentSize = 0;
+    void *buffer2 = NULL;
+    size_t buffer2Size = 0;
+    size_t buffer2ContentSize = 0;
+    void *destBuffer = NULL;
+    PyObject *result = NULL;
+    ZSTD_outBuffer outBuffer;
+    ZSTD_inBuffer inBuffer;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs,
+                                     "O!:decompress_content_dict_chain", kwlist,
+                                     &PyList_Type, &chunks)) {
+        return NULL;
+    }
+
+    chunksLen = PyList_Size(chunks);
+    if (!chunksLen) {
+        PyErr_SetString(PyExc_ValueError, "empty input chain");
+        return NULL;
+    }
+
+    /* The first chunk should not be using a dictionary. We handle it specially.
+     */
+    chunk = PyList_GetItem(chunks, 0);
+    if (!PyBytes_Check(chunk)) {
+        PyErr_SetString(PyExc_ValueError, "chunk 0 must be bytes");
+        return NULL;
+    }
+
+    /* We require that all chunks be zstd frames and that they have content size
+     * set. */
+    PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
+    zresult = ZSTD_getFrameHeader(&frameHeader, (void *)chunkData, chunkSize);
+    if (ZSTD_isError(zresult)) {
+        PyErr_SetString(PyExc_ValueError, "chunk 0 is not a valid zstd frame");
+        return NULL;
+    }
+    else if (zresult) {
+        PyErr_SetString(PyExc_ValueError,
+                        "chunk 0 is too small to contain a zstd frame");
+        return NULL;
+    }
+
+    if (ZSTD_CONTENTSIZE_UNKNOWN == frameHeader.frameContentSize) {
+        PyErr_SetString(PyExc_ValueError,
+                        "chunk 0 missing content size in frame");
+        return NULL;
+    }
+
+    assert(ZSTD_CONTENTSIZE_ERROR != frameHeader.frameContentSize);
+
+    /* We check against PY_SSIZE_T_MAX here because we ultimately cast the
+     * result to a Python object and it's length can be no greater than
+     * Py_ssize_t. In theory, we could have an intermediate frame that is
+     * larger. But a) why would this API be used for frames that large b)
+     * it isn't worth the complexity to support. */
+    assert(SIZE_MAX >= PY_SSIZE_T_MAX);
+    if (frameHeader.frameContentSize > PY_SSIZE_T_MAX) {
+        PyErr_SetString(PyExc_ValueError,
+                        "chunk 0 is too large to decompress on this platform");
+        return NULL;
+    }
+
+    if (ensure_dctx(self, 0)) {
+        goto finally;
+    }
+
+    buffer1Size = (size_t)frameHeader.frameContentSize;
+    buffer1 = PyMem_Malloc(buffer1Size);
+    if (!buffer1) {
+        goto finally;
+    }
+
+    outBuffer.dst = buffer1;
+    outBuffer.size = buffer1Size;
+    outBuffer.pos = 0;
+
+    inBuffer.src = chunkData;
+    inBuffer.size = chunkSize;
+    inBuffer.pos = 0;
+
+    Py_BEGIN_ALLOW_THREADS zresult =
+        ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+    Py_END_ALLOW_THREADS if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "could not decompress chunk 0: %s",
+                     ZSTD_getErrorName(zresult));
+        goto finally;
+    }
+    else if (zresult) {
+        PyErr_Format(ZstdError, "chunk 0 did not decompress full frame");
+        goto finally;
+    }
+
+    buffer1ContentSize = outBuffer.pos;
+
+    /* Special case of a simple chain. */
+    if (1 == chunksLen) {
+        result = PyBytes_FromStringAndSize(buffer1, buffer1Size);
+        goto finally;
+    }
+
+    /* This should ideally look at next chunk. But this is slightly simpler. */
+    buffer2Size = (size_t)frameHeader.frameContentSize;
+    buffer2 = PyMem_Malloc(buffer2Size);
+    if (!buffer2) {
+        goto finally;
+    }
+
+    /* For each subsequent chunk, use the previous fulltext as a content
+       dictionary. Our strategy is to have 2 buffers. One holds the previous
+       fulltext (to be used as a content dictionary) and the other holds the new
+       fulltext. The buffers grow when needed but never decrease in size. This
+       limits the memory allocator overhead.
+    */
+    for (chunkIndex = 1; chunkIndex < chunksLen; chunkIndex++) {
+        chunk = PyList_GetItem(chunks, chunkIndex);
+        if (!PyBytes_Check(chunk)) {
+            PyErr_Format(PyExc_ValueError, "chunk %zd must be bytes",
+                         chunkIndex);
+            goto finally;
+        }
+
+        PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
+        zresult =
+            ZSTD_getFrameHeader(&frameHeader, (void *)chunkData, chunkSize);
+        if (ZSTD_isError(zresult)) {
+            PyErr_Format(PyExc_ValueError,
+                         "chunk %zd is not a valid zstd frame", chunkIndex);
+            goto finally;
+        }
+        else if (zresult) {
+            PyErr_Format(PyExc_ValueError,
+                         "chunk %zd is too small to contain a zstd frame",
+                         chunkIndex);
+            goto finally;
+        }
+
+        if (ZSTD_CONTENTSIZE_UNKNOWN == frameHeader.frameContentSize) {
+            PyErr_Format(PyExc_ValueError,
+                         "chunk %zd missing content size in frame", chunkIndex);
+            goto finally;
+        }
+
+        assert(ZSTD_CONTENTSIZE_ERROR != frameHeader.frameContentSize);
+
+        if (frameHeader.frameContentSize > PY_SSIZE_T_MAX) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "chunk %zd is too large to decompress on this platform",
+                chunkIndex);
+            goto finally;
+        }
+
+        inBuffer.src = chunkData;
+        inBuffer.size = chunkSize;
+        inBuffer.pos = 0;
+
+        parity = chunkIndex % 2;
+
+        /* This could definitely be abstracted to reduce code duplication. */
+        if (parity) {
+            /* Resize destination buffer to hold larger content. */
+            if (buffer2Size < frameHeader.frameContentSize) {
+                buffer2Size = (size_t)frameHeader.frameContentSize;
+                destBuffer = PyMem_Realloc(buffer2, buffer2Size);
+                if (!destBuffer) {
+                    goto finally;
+                }
+                buffer2 = destBuffer;
+            }
+
+            Py_BEGIN_ALLOW_THREADS zresult = ZSTD_DCtx_refPrefix_advanced(
+                self->dctx, buffer1, buffer1ContentSize, ZSTD_dct_rawContent);
+            Py_END_ALLOW_THREADS if (ZSTD_isError(zresult)) {
+                PyErr_Format(ZstdError,
+                             "failed to load prefix dictionary at chunk %zd",
+                             chunkIndex);
+                goto finally;
+            }
+
+            outBuffer.dst = buffer2;
+            outBuffer.size = buffer2Size;
+            outBuffer.pos = 0;
+
+            Py_BEGIN_ALLOW_THREADS zresult =
+                ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+            Py_END_ALLOW_THREADS if (ZSTD_isError(zresult)) {
+                PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
+                             chunkIndex, ZSTD_getErrorName(zresult));
+                goto finally;
+            }
+            else if (zresult) {
+                PyErr_Format(ZstdError,
+                             "chunk %zd did not decompress full frame",
+                             chunkIndex);
+                goto finally;
+            }
+
+            buffer2ContentSize = outBuffer.pos;
+        }
+        else {
+            if (buffer1Size < frameHeader.frameContentSize) {
+                buffer1Size = (size_t)frameHeader.frameContentSize;
+                destBuffer = PyMem_Realloc(buffer1, buffer1Size);
+                if (!destBuffer) {
+                    goto finally;
+                }
+                buffer1 = destBuffer;
+            }
+
+            Py_BEGIN_ALLOW_THREADS zresult = ZSTD_DCtx_refPrefix_advanced(
+                self->dctx, buffer2, buffer2ContentSize, ZSTD_dct_rawContent);
+            Py_END_ALLOW_THREADS if (ZSTD_isError(zresult)) {
+                PyErr_Format(ZstdError,
+                             "failed to load prefix dictionary at chunk %zd",
+                             chunkIndex);
+                goto finally;
+            }
+
+            outBuffer.dst = buffer1;
+            outBuffer.size = buffer1Size;
+            outBuffer.pos = 0;
+
+            Py_BEGIN_ALLOW_THREADS zresult =
+                ZSTD_decompressStream(self->dctx, &outBuffer, &inBuffer);
+            Py_END_ALLOW_THREADS if (ZSTD_isError(zresult)) {
+                PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
+                             chunkIndex, ZSTD_getErrorName(zresult));
+                goto finally;
+            }
+            else if (zresult) {
+                PyErr_Format(ZstdError,
+                             "chunk %zd did not decompress full frame",
+                             chunkIndex);
+                goto finally;
+            }
+
+            buffer1ContentSize = outBuffer.pos;
+        }
+    }
+
+    result = PyBytes_FromStringAndSize(parity ? buffer2 : buffer1,
+                                       parity ? buffer2ContentSize
+                                              : buffer1ContentSize);
+
+finally:
+    if (buffer2) {
+        PyMem_Free(buffer2);
+    }
+    if (buffer1) {
+        PyMem_Free(buffer1);
+    }
+
+    return result;
+}
+
+typedef struct {
+    void *sourceData;
+    size_t sourceSize;
+    size_t destSize;
+} FramePointer;
+
+typedef struct {
+    FramePointer *frames;
+    Py_ssize_t framesSize;
+    unsigned long long compressedSize;
+} FrameSources;
+
+typedef struct {
+    void *dest;
+    Py_ssize_t destSize;
+    BufferSegment *segments;
+    Py_ssize_t segmentsSize;
+} DecompressorDestBuffer;
+
+typedef enum {
+    DecompressorWorkerError_none = 0,
+    DecompressorWorkerError_zstd = 1,
+    DecompressorWorkerError_memory = 2,
+    DecompressorWorkerError_sizeMismatch = 3,
+    DecompressorWorkerError_unknownSize = 4,
+} DecompressorWorkerError;
+
+typedef struct {
+    /* Source records and length */
+    FramePointer *framePointers;
+    /* Which records to process. */
+    Py_ssize_t startOffset;
+    Py_ssize_t endOffset;
+    unsigned long long totalSourceSize;
+
+    /* Compression state and settings. */
+    ZSTD_DCtx *dctx;
+    int requireOutputSizes;
+
+    /* Output storage. */
+    DecompressorDestBuffer *destBuffers;
+    Py_ssize_t destCount;
+
+    /* Item that error occurred on. */
+    Py_ssize_t errorOffset;
+    /* If an error occurred. */
+    DecompressorWorkerError error;
+    /* result from zstd decompression operation */
+    size_t zresult;
+} DecompressorWorkerState;
+
+#ifdef HAVE_ZSTD_POOL_APIS
+static void decompress_worker(DecompressorWorkerState *state) {
+    size_t allocationSize;
+    DecompressorDestBuffer *destBuffer;
+    Py_ssize_t frameIndex;
+    Py_ssize_t localOffset = 0;
+    Py_ssize_t currentBufferStartIndex = state->startOffset;
+    Py_ssize_t remainingItems = state->endOffset - state->startOffset + 1;
+    void *tmpBuf;
+    Py_ssize_t destOffset = 0;
+    FramePointer *framePointers = state->framePointers;
+    size_t zresult;
+
+    assert(NULL == state->destBuffers);
+    assert(0 == state->destCount);
+    assert(state->endOffset - state->startOffset >= 0);
+
+    /* We could get here due to the way work is allocated. Ideally we wouldn't
+       get here. But that would require a bit of a refactor in the caller. */
+    if (state->totalSourceSize > SIZE_MAX) {
+        state->error = DecompressorWorkerError_memory;
+        state->errorOffset = 0;
+        return;
+    }
+
+    /*
+     * We need to allocate a buffer to hold decompressed data. How we do this
+     * depends on what we know about the output. The following scenarios are
+     * possible:
+     *
+     * 1. All structs defining frames declare the output size.
+     * 2. The decompressed size is embedded within the zstd frame.
+     * 3. The decompressed size is not stored anywhere.
+     *
+     * For now, we only support #1 and #2.
+     */
+
+    /* Resolve ouput segments. */
+    for (frameIndex = state->startOffset; frameIndex <= state->endOffset;
+         frameIndex++) {
+        FramePointer *fp = &framePointers[frameIndex];
+        unsigned long long decompressedSize;
+
+        if (0 == fp->destSize) {
+            decompressedSize =
+                ZSTD_getFrameContentSize(fp->sourceData, fp->sourceSize);
+
+            if (ZSTD_CONTENTSIZE_ERROR == decompressedSize) {
+                state->error = DecompressorWorkerError_unknownSize;
+                state->errorOffset = frameIndex;
+                return;
+            }
+            else if (ZSTD_CONTENTSIZE_UNKNOWN == decompressedSize) {
+                if (state->requireOutputSizes) {
+                    state->error = DecompressorWorkerError_unknownSize;
+                    state->errorOffset = frameIndex;
+                    return;
+                }
+
+                /* This will fail the assert for .destSize > 0 below. */
+                decompressedSize = 0;
+            }
+
+            if (decompressedSize > SIZE_MAX) {
+                state->error = DecompressorWorkerError_memory;
+                state->errorOffset = frameIndex;
+                return;
+            }
+
+            fp->destSize = (size_t)decompressedSize;
+        }
+    }
+
+    state->destBuffers = calloc(1, sizeof(DecompressorDestBuffer));
+    if (NULL == state->destBuffers) {
+        state->error = DecompressorWorkerError_memory;
+        return;
+    }
+
+    state->destCount = 1;
+
+    destBuffer = &state->destBuffers[state->destCount - 1];
+
+    assert(framePointers[state->startOffset].destSize > 0); /* For now. */
+
+    allocationSize = roundpow2((size_t)state->totalSourceSize);
+
+    if (framePointers[state->startOffset].destSize > allocationSize) {
+        allocationSize = roundpow2(framePointers[state->startOffset].destSize);
+    }
+
+    destBuffer->dest = malloc(allocationSize);
+    if (NULL == destBuffer->dest) {
+        state->error = DecompressorWorkerError_memory;
+        return;
+    }
+
+    destBuffer->destSize = allocationSize;
+
+    destBuffer->segments = calloc(remainingItems, sizeof(BufferSegment));
+    if (NULL == destBuffer->segments) {
+        /* Caller will free state->dest as part of cleanup. */
+        state->error = DecompressorWorkerError_memory;
+        return;
+    }
+
+    destBuffer->segmentsSize = remainingItems;
+
+    for (frameIndex = state->startOffset; frameIndex <= state->endOffset;
+         frameIndex++) {
+        ZSTD_outBuffer outBuffer;
+        ZSTD_inBuffer inBuffer;
+        const void *source = framePointers[frameIndex].sourceData;
+        const size_t sourceSize = framePointers[frameIndex].sourceSize;
+        void *dest;
+        const size_t decompressedSize = framePointers[frameIndex].destSize;
+        size_t destAvailable = destBuffer->destSize - destOffset;
+
+        assert(decompressedSize > 0); /* For now. */
+
+        /*
+         * Not enough space in current buffer. Finish current before and
+         * allocate and switch to a new one.
+         */
+        if (decompressedSize > destAvailable) {
+            /*
+             * Shrinking the destination buffer is optional. But it should be
+             * cheap, so we just do it.
+             */
+            if (destAvailable) {
+                tmpBuf = realloc(destBuffer->dest, destOffset);
+                if (NULL == tmpBuf) {
+                    state->error = DecompressorWorkerError_memory;
+                    return;
+                }
+
+                destBuffer->dest = tmpBuf;
+                destBuffer->destSize = destOffset;
+            }
+
+            /* Truncate segments buffer. */
+            tmpBuf = realloc(destBuffer->segments,
+                             (frameIndex - currentBufferStartIndex) *
+                                 sizeof(BufferSegment));
+            if (NULL == tmpBuf) {
+                state->error = DecompressorWorkerError_memory;
+                return;
+            }
+
+            destBuffer->segments = tmpBuf;
+            destBuffer->segmentsSize = frameIndex - currentBufferStartIndex;
+
+            /* Grow space for new DestBuffer. */
+            tmpBuf =
+                realloc(state->destBuffers, (state->destCount + 1) *
+                                                sizeof(DecompressorDestBuffer));
+            if (NULL == tmpBuf) {
+                state->error = DecompressorWorkerError_memory;
+                return;
+            }
+
+            state->destBuffers = tmpBuf;
+            state->destCount++;
+
+            destBuffer = &state->destBuffers[state->destCount - 1];
+
+            /* Don't take any chances will non-NULL pointers. */
+            memset(destBuffer, 0, sizeof(DecompressorDestBuffer));
+
+            allocationSize = roundpow2((size_t)state->totalSourceSize);
+
+            if (decompressedSize > allocationSize) {
+                allocationSize = roundpow2(decompressedSize);
+            }
+
+            destBuffer->dest = malloc(allocationSize);
+            if (NULL == destBuffer->dest) {
+                state->error = DecompressorWorkerError_memory;
+                return;
+            }
+
+            destBuffer->destSize = allocationSize;
+            destAvailable = allocationSize;
+            destOffset = 0;
+            localOffset = 0;
+
+            destBuffer->segments =
+                calloc(remainingItems, sizeof(BufferSegment));
+            if (NULL == destBuffer->segments) {
+                state->error = DecompressorWorkerError_memory;
+                return;
+            }
+
+            destBuffer->segmentsSize = remainingItems;
+            currentBufferStartIndex = frameIndex;
+        }
+
+        dest = (char *)destBuffer->dest + destOffset;
+
+        outBuffer.dst = dest;
+        outBuffer.size = decompressedSize;
+        outBuffer.pos = 0;
+
+        inBuffer.src = source;
+        inBuffer.size = sourceSize;
+        inBuffer.pos = 0;
+
+        zresult = ZSTD_decompressStream(state->dctx, &outBuffer, &inBuffer);
+        if (ZSTD_isError(zresult)) {
+            state->error = DecompressorWorkerError_zstd;
+            state->zresult = zresult;
+            state->errorOffset = frameIndex;
+            return;
+        }
+        else if (zresult || outBuffer.pos != decompressedSize) {
+            state->error = DecompressorWorkerError_sizeMismatch;
+            state->zresult = outBuffer.pos;
+            state->errorOffset = frameIndex;
+            return;
+        }
+
+        destBuffer->segments[localOffset].offset = destOffset;
+        destBuffer->segments[localOffset].length = outBuffer.pos;
+        destOffset += outBuffer.pos;
+        localOffset++;
+        remainingItems--;
+    }
+
+    if (destBuffer->destSize > destOffset) {
+        tmpBuf = realloc(destBuffer->dest, destOffset);
+        if (NULL == tmpBuf) {
+            state->error = DecompressorWorkerError_memory;
+            return;
+        }
+
+        destBuffer->dest = tmpBuf;
+        destBuffer->destSize = destOffset;
+    }
+}
+#endif
+
+#ifdef HAVE_ZSTD_POOL_APIS
+ZstdBufferWithSegmentsCollection *
+decompress_from_framesources(ZstdDecompressor *decompressor,
+                             FrameSources *frames, Py_ssize_t threadCount) {
+    Py_ssize_t i = 0;
+    int errored = 0;
+    Py_ssize_t segmentsCount;
+    ZstdBufferWithSegments *bws = NULL;
+    PyObject *resultArg = NULL;
+    Py_ssize_t resultIndex;
+    ZstdBufferWithSegmentsCollection *result = NULL;
+    FramePointer *framePointers = frames->frames;
+    unsigned long long workerBytes = 0;
+    Py_ssize_t currentThread = 0;
+    Py_ssize_t workerStartOffset = 0;
+    POOL_ctx *pool = NULL;
+    DecompressorWorkerState *workerStates = NULL;
+    unsigned long long bytesPerWorker;
+
+    /* Caller should normalize 0 and negative values to 1 or larger. */
+    assert(threadCount >= 1);
+
+    /* More threads than inputs makes no sense under any conditions. */
+    threadCount =
+        frames->framesSize < threadCount ? frames->framesSize : threadCount;
+
+    /* TODO lower thread count if input size is too small and threads would just
+       add overhead. */
+
+    if (decompressor->dict) {
+        if (ensure_ddict(decompressor->dict)) {
+            return NULL;
+        }
+    }
+
+    /* If threadCount==1, we don't start a thread pool. But we do leverage the
+       same API for dispatching work. */
+    workerStates = PyMem_Malloc(threadCount * sizeof(DecompressorWorkerState));
+    if (NULL == workerStates) {
+        PyErr_NoMemory();
+        goto finally;
+    }
+
+    memset(workerStates, 0, threadCount * sizeof(DecompressorWorkerState));
+
+    if (threadCount > 1) {
+        pool = POOL_create(threadCount, 1);
+        if (NULL == pool) {
+            PyErr_SetString(ZstdError, "could not initialize zstd thread pool");
+            goto finally;
+        }
+    }
+
+    bytesPerWorker = frames->compressedSize / threadCount;
+
+    if (bytesPerWorker > SIZE_MAX) {
+        PyErr_SetString(ZstdError,
+                        "too much data per worker for this platform");
+        goto finally;
+    }
+
+    for (i = 0; i < threadCount; i++) {
+        size_t zresult;
+
+        workerStates[i].dctx = ZSTD_createDCtx();
+        if (NULL == workerStates[i].dctx) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        if (decompressor->dict) {
+            zresult = ZSTD_DCtx_refDDict(workerStates[i].dctx,
+                                         decompressor->dict->ddict);
+            if (zresult) {
+                PyErr_Format(ZstdError,
+                             "unable to reference prepared dictionary: %s",
+                             ZSTD_getErrorName(zresult));
+                goto finally;
+            }
+        }
+
+        workerStates[i].framePointers = framePointers;
+        workerStates[i].requireOutputSizes = 1;
+    }
+
+    Py_BEGIN_ALLOW_THREADS
+        /* There are many ways to split work among workers.
+
+           For now, we take a simple approach of splitting work so each worker
+           gets roughly the same number of input bytes. This will result in more
+           starvation than running N>threadCount jobs. But it avoids
+           complications around state tracking, which could involve extra
+           locking.
+        */
+        for (i = 0; i < frames->framesSize; i++) {
+        workerBytes += frames->frames[i].sourceSize;
+
+        /*
+         * The last worker/thread needs to handle all remaining work. Don't
+         * trigger it prematurely. Defer to the block outside of the loop.
+         * (But still process this loop so workerBytes is correct.
+         */
+        if (currentThread == threadCount - 1) {
+            continue;
+        }
+
+        if (workerBytes >= bytesPerWorker) {
+            workerStates[currentThread].startOffset = workerStartOffset;
+            workerStates[currentThread].endOffset = i;
+            workerStates[currentThread].totalSourceSize = workerBytes;
+
+            if (threadCount > 1) {
+                POOL_add(pool, (POOL_function)decompress_worker,
+                         &workerStates[currentThread]);
+            }
+            else {
+                decompress_worker(&workerStates[currentThread]);
+            }
+            currentThread++;
+            workerStartOffset = i + 1;
+            workerBytes = 0;
+        }
+    }
+
+    if (workerBytes) {
+        workerStates[currentThread].startOffset = workerStartOffset;
+        workerStates[currentThread].endOffset = frames->framesSize - 1;
+        workerStates[currentThread].totalSourceSize = workerBytes;
+
+        if (threadCount > 1) {
+            POOL_add(pool, (POOL_function)decompress_worker,
+                     &workerStates[currentThread]);
+        }
+        else {
+            decompress_worker(&workerStates[currentThread]);
+        }
+    }
+
+    if (threadCount > 1) {
+        POOL_free(pool);
+        pool = NULL;
+    }
+    Py_END_ALLOW_THREADS
+
+        for (i = 0; i < threadCount; i++) {
+        switch (workerStates[i].error) {
+        case DecompressorWorkerError_none:
+            break;
+
+        case DecompressorWorkerError_zstd:
+            PyErr_Format(ZstdError, "error decompressing item %zd: %s",
+                         workerStates[i].errorOffset,
+                         ZSTD_getErrorName(workerStates[i].zresult));
+            errored = 1;
+            break;
+
+        case DecompressorWorkerError_memory:
+            PyErr_NoMemory();
+            errored = 1;
+            break;
+
+        case DecompressorWorkerError_sizeMismatch:
+            PyErr_Format(ZstdError,
+                         "error decompressing item %zd: decompressed %zu "
+                         "bytes; expected %zu",
+                         workerStates[i].errorOffset, workerStates[i].zresult,
+                         framePointers[workerStates[i].errorOffset].destSize);
+            errored = 1;
+            break;
+
+        case DecompressorWorkerError_unknownSize:
+            PyErr_Format(PyExc_ValueError,
+                         "could not determine decompressed size of item %zd",
+                         workerStates[i].errorOffset);
+            errored = 1;
+            break;
+
+        default:
+            PyErr_Format(ZstdError, "unhandled error type: %d; this is a bug",
+                         workerStates[i].error);
+            errored = 1;
+            break;
+        }
+
+        if (errored) {
+            break;
+        }
+    }
+
+    if (errored) {
+        goto finally;
+    }
+
+    segmentsCount = 0;
+    for (i = 0; i < threadCount; i++) {
+        segmentsCount += workerStates[i].destCount;
+    }
+
+    resultArg = PyTuple_New(segmentsCount);
+    if (NULL == resultArg) {
+        goto finally;
+    }
+
+    resultIndex = 0;
+
+    for (i = 0; i < threadCount; i++) {
+        Py_ssize_t bufferIndex;
+        DecompressorWorkerState *state = &workerStates[i];
+
+        for (bufferIndex = 0; bufferIndex < state->destCount; bufferIndex++) {
+            DecompressorDestBuffer *destBuffer =
+                &state->destBuffers[bufferIndex];
+
+            bws = BufferWithSegments_FromMemory(
+                destBuffer->dest, destBuffer->destSize, destBuffer->segments,
+                destBuffer->segmentsSize);
+            if (NULL == bws) {
+                goto finally;
+            }
+
+            /*
+             * Memory for buffer and segments was allocated using malloc() in
+             * worker and the memory is transferred to the BufferWithSegments
+             * instance. So tell instance to use free() and NULL the reference
+             * in the state struct so it isn't freed below.
+             */
+            bws->useFree = 1;
+            destBuffer->dest = NULL;
+            destBuffer->segments = NULL;
+
+            PyTuple_SET_ITEM(resultArg, resultIndex++, (PyObject *)bws);
+        }
+    }
+
+    result = (ZstdBufferWithSegmentsCollection *)PyObject_CallObject(
+        (PyObject *)ZstdBufferWithSegmentsCollectionType, resultArg);
+
+finally:
+    Py_CLEAR(resultArg);
+
+    if (workerStates) {
+        for (i = 0; i < threadCount; i++) {
+            Py_ssize_t bufferIndex;
+            DecompressorWorkerState *state = &workerStates[i];
+
+            if (state->dctx) {
+                ZSTD_freeDCtx(state->dctx);
+            }
+
+            for (bufferIndex = 0; bufferIndex < state->destCount;
+                 bufferIndex++) {
+                if (state->destBuffers) {
+                    /*
+                     * Will be NULL if memory transfered to a
+                     * BufferWithSegments. Otherwise it is left over after an
+                     * error occurred.
+                     */
+                    free(state->destBuffers[bufferIndex].dest);
+                    free(state->destBuffers[bufferIndex].segments);
+                }
+            }
+
+            free(state->destBuffers);
+        }
+
+        PyMem_Free(workerStates);
+    }
+
+    POOL_free(pool);
+
+    return result;
+}
+#endif
+
+#ifdef HAVE_ZSTD_POOL_APIS
+static ZstdBufferWithSegmentsCollection *
+Decompressor_multi_decompress_to_buffer(ZstdDecompressor *self, PyObject *args,
+                                        PyObject *kwargs) {
+    static char *kwlist[] = {"frames", "decompressed_sizes", "threads", NULL};
+
+    PyObject *frames;
+    Py_buffer frameSizes;
+    int threads = 0;
+    Py_ssize_t frameCount;
+    Py_buffer *frameBuffers = NULL;
+    FramePointer *framePointers = NULL;
+    unsigned long long *frameSizesP = NULL;
+    unsigned long long totalInputSize = 0;
+    FrameSources frameSources;
+    ZstdBufferWithSegmentsCollection *result = NULL;
+    Py_ssize_t i;
+
+    memset(&frameSizes, 0, sizeof(frameSizes));
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs,
+                                     "O|y*i:multi_decompress_to_buffer", kwlist,
+                                     &frames, &frameSizes, &threads)) {
+        return NULL;
+    }
+
+    if (frameSizes.buf) {
+        frameSizesP = (unsigned long long *)frameSizes.buf;
+    }
+
+    if (threads < 0) {
+        threads = cpu_count();
+    }
+
+    if (threads < 2) {
+        threads = 1;
+    }
+
+    if (PyObject_TypeCheck(frames, ZstdBufferWithSegmentsType)) {
+        ZstdBufferWithSegments *buffer = (ZstdBufferWithSegments *)frames;
+        frameCount = buffer->segmentCount;
+
+        if (frameSizes.buf &&
+            frameSizes.len !=
+                frameCount * (Py_ssize_t)sizeof(unsigned long long)) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "decompressed_sizes size mismatch; expected %zd, got %zd",
+                frameCount * sizeof(unsigned long long), frameSizes.len);
+            goto finally;
+        }
+
+        framePointers = PyMem_Malloc(frameCount * sizeof(FramePointer));
+        if (!framePointers) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        for (i = 0; i < frameCount; i++) {
+            void *sourceData;
+            unsigned long long sourceSize;
+            unsigned long long decompressedSize = 0;
+
+            if (buffer->segments[i].offset + buffer->segments[i].length >
+                buffer->dataSize) {
+                PyErr_Format(PyExc_ValueError,
+                             "item %zd has offset outside memory area", i);
+                goto finally;
+            }
+
+            sourceData = (char *)buffer->data + buffer->segments[i].offset;
+            sourceSize = buffer->segments[i].length;
+            totalInputSize += sourceSize;
+
+            if (frameSizesP) {
+                decompressedSize = frameSizesP[i];
+            }
+
+            if (sourceSize > SIZE_MAX) {
+                PyErr_Format(PyExc_ValueError,
+                             "item %zd is too large for this platform", i);
+                goto finally;
+            }
+
+            if (decompressedSize > SIZE_MAX) {
+                PyErr_Format(PyExc_ValueError,
+                             "decompressed size of item %zd is too large for "
+                             "this platform",
+                             i);
+                goto finally;
+            }
+
+            framePointers[i].sourceData = sourceData;
+            framePointers[i].sourceSize = (size_t)sourceSize;
+            framePointers[i].destSize = (size_t)decompressedSize;
+        }
+    }
+    else if (PyObject_TypeCheck(frames, ZstdBufferWithSegmentsCollectionType)) {
+        Py_ssize_t offset = 0;
+        ZstdBufferWithSegments *buffer;
+        ZstdBufferWithSegmentsCollection *collection =
+            (ZstdBufferWithSegmentsCollection *)frames;
+
+        frameCount = BufferWithSegmentsCollection_length(collection);
+
+        if (frameSizes.buf && frameSizes.len != frameCount) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "decompressed_sizes size mismatch; expected %zd; got %zd",
+                frameCount * sizeof(unsigned long long), frameSizes.len);
+            goto finally;
+        }
+
+        framePointers = PyMem_Malloc(frameCount * sizeof(FramePointer));
+        if (NULL == framePointers) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        /* Iterate the data structure directly because it is faster. */
+        for (i = 0; i < collection->bufferCount; i++) {
+            Py_ssize_t segmentIndex;
+            buffer = collection->buffers[i];
+
+            for (segmentIndex = 0; segmentIndex < buffer->segmentCount;
+                 segmentIndex++) {
+                unsigned long long decompressedSize =
+                    frameSizesP ? frameSizesP[offset] : 0;
+
+                if (buffer->segments[segmentIndex].offset +
+                        buffer->segments[segmentIndex].length >
+                    buffer->dataSize) {
+                    PyErr_Format(PyExc_ValueError,
+                                 "item %zd has offset outside memory area",
+                                 offset);
+                    goto finally;
+                }
+
+                if (buffer->segments[segmentIndex].length > SIZE_MAX) {
+                    PyErr_Format(
+                        PyExc_ValueError,
+                        "item %zd in buffer %zd is too large for this platform",
+                        segmentIndex, i);
+                    goto finally;
+                }
+
+                if (decompressedSize > SIZE_MAX) {
+                    PyErr_Format(PyExc_ValueError,
+                                 "decompressed size of item %zd in buffer %zd "
+                                 "is too large for this platform",
+                                 segmentIndex, i);
+                    goto finally;
+                }
+
+                totalInputSize += buffer->segments[segmentIndex].length;
+
+                framePointers[offset].sourceData =
+                    (char *)buffer->data +
+                    buffer->segments[segmentIndex].offset;
+                framePointers[offset].sourceSize =
+                    (size_t)buffer->segments[segmentIndex].length;
+                framePointers[offset].destSize = (size_t)decompressedSize;
+
+                offset++;
+            }
+        }
+    }
+    else if (PyList_Check(frames)) {
+        frameCount = PyList_GET_SIZE(frames);
+
+        if (frameSizes.buf &&
+            frameSizes.len !=
+                frameCount * (Py_ssize_t)sizeof(unsigned long long)) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "decompressed_sizes size mismatch; expected %zd, got %zd",
+                frameCount * sizeof(unsigned long long), frameSizes.len);
+            goto finally;
+        }
+
+        framePointers = PyMem_Malloc(frameCount * sizeof(FramePointer));
+        if (!framePointers) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        frameBuffers = PyMem_Malloc(frameCount * sizeof(Py_buffer));
+        if (NULL == frameBuffers) {
+            PyErr_NoMemory();
+            goto finally;
+        }
+
+        memset(frameBuffers, 0, frameCount * sizeof(Py_buffer));
+
+        /* Do a pass to assemble info about our input buffers and output sizes.
+         */
+        for (i = 0; i < frameCount; i++) {
+            unsigned long long decompressedSize =
+                frameSizesP ? frameSizesP[i] : 0;
+
+            if (0 != PyObject_GetBuffer(PyList_GET_ITEM(frames, i),
+                                        &frameBuffers[i], PyBUF_CONTIG_RO)) {
+                PyErr_Clear();
+                PyErr_Format(PyExc_TypeError,
+                             "item %zd not a bytes like object", i);
+                goto finally;
+            }
+
+            if (decompressedSize > SIZE_MAX) {
+                PyErr_Format(PyExc_ValueError,
+                             "decompressed size of item %zd is too large for "
+                             "this platform",
+                             i);
+                goto finally;
+            }
+
+            totalInputSize += frameBuffers[i].len;
+
+            framePointers[i].sourceData = frameBuffers[i].buf;
+            framePointers[i].sourceSize = frameBuffers[i].len;
+            framePointers[i].destSize = (size_t)decompressedSize;
+        }
+    }
+    else {
+        PyErr_SetString(PyExc_TypeError,
+                        "argument must be list or BufferWithSegments");
+        goto finally;
+    }
+
+    /* We now have an array with info about our inputs and outputs. Feed it into
+       our generic decompression function. */
+    frameSources.frames = framePointers;
+    frameSources.framesSize = frameCount;
+    frameSources.compressedSize = totalInputSize;
+
+    result = decompress_from_framesources(self, &frameSources, threads);
+
+finally:
+    if (frameSizes.buf) {
+        PyBuffer_Release(&frameSizes);
+    }
+    PyMem_Free(framePointers);
+
+    if (frameBuffers) {
+        for (i = 0; i < frameCount; i++) {
+            PyBuffer_Release(&frameBuffers[i]);
+        }
+
+        PyMem_Free(frameBuffers);
+    }
+
+    return result;
+}
+#endif
+
+static PyMethodDef Decompressor_methods[] = {
+    {"copy_stream", (PyCFunction)Decompressor_copy_stream,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"decompress", (PyCFunction)Decompressor_decompress,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"decompressobj", (PyCFunction)Decompressor_decompressobj,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"read_to_iter", (PyCFunction)Decompressor_read_to_iter,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"stream_reader", (PyCFunction)Decompressor_stream_reader,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"stream_writer", (PyCFunction)Decompressor_stream_writer,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"decompress_content_dict_chain",
+     (PyCFunction)Decompressor_decompress_content_dict_chain,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+#ifdef HAVE_ZSTD_POOL_APIS
+    {"multi_decompress_to_buffer",
+     (PyCFunction)Decompressor_multi_decompress_to_buffer,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+#endif
+    {"memory_size", (PyCFunction)Decompressor_memory_size, METH_NOARGS, NULL},
+    {NULL, NULL}};
+
+PyType_Slot ZstdDecompressorSlots[] = {
+    {Py_tp_dealloc, Decompressor_dealloc},
+    {Py_tp_methods, Decompressor_methods},
+    {Py_tp_init, Decompressor_init},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdDecompressorSpec = {
+    "zstd.ZstdDecompressor",
+    sizeof(ZstdDecompressor),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdDecompressorSlots,
+};
+
+PyTypeObject *ZstdDecompressorType;
+
+void decompressor_module_init(PyObject *mod) {
+    ZstdDecompressorType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdDecompressorSpec);
+    if (PyType_Ready(ZstdDecompressorType) < 0) {
+        return;
+    }
+
+    Py_INCREF((PyObject *)ZstdDecompressorType);
+    PyModule_AddObject(mod, "ZstdDecompressor",
+                       (PyObject *)ZstdDecompressorType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/decompressoriterator.c b/contrib/python/zstandard/py3/c-ext/decompressoriterator.c
new file mode 100644
index 00000000000..e575a7a1016
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/decompressoriterator.c
@@ -0,0 +1,228 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+#define min(a, b) (((a) < (b)) ? (a) : (b))
+
+extern PyObject *ZstdError;
+
+static void ZstdDecompressorIterator_dealloc(ZstdDecompressorIterator *self) {
+    Py_XDECREF(self->decompressor);
+    Py_XDECREF(self->reader);
+
+    if (self->buffer.buf) {
+        PyBuffer_Release(&self->buffer);
+        memset(&self->buffer, 0, sizeof(self->buffer));
+    }
+
+    if (self->input.src) {
+        PyMem_Free((void *)self->input.src);
+        self->input.src = NULL;
+    }
+
+    PyObject_Del(self);
+}
+
+static PyObject *ZstdDecompressorIterator_iter(PyObject *self) {
+    Py_INCREF(self);
+    return self;
+}
+
+static DecompressorIteratorResult
+read_decompressor_iterator(ZstdDecompressorIterator *self) {
+    size_t zresult;
+    PyObject *chunk;
+    DecompressorIteratorResult result;
+    size_t oldInputPos = self->input.pos;
+
+    result.chunk = NULL;
+
+    chunk = PyBytes_FromStringAndSize(NULL, self->outSize);
+    if (!chunk) {
+        result.errored = 1;
+        return result;
+    }
+
+    self->output.dst = PyBytes_AsString(chunk);
+    self->output.size = self->outSize;
+    self->output.pos = 0;
+
+    Py_BEGIN_ALLOW_THREADS zresult = ZSTD_decompressStream(
+        self->decompressor->dctx, &self->output, &self->input);
+    Py_END_ALLOW_THREADS
+
+        /* We're done with the pointer. Nullify to prevent anyone from getting a
+        handle on a Python object. */
+        self->output.dst = NULL;
+
+    if (ZSTD_isError(zresult)) {
+        Py_DECREF(chunk);
+        PyErr_Format(ZstdError, "zstd decompress error: %s",
+                     ZSTD_getErrorName(zresult));
+        result.errored = 1;
+        return result;
+    }
+
+    self->readCount += self->input.pos - oldInputPos;
+
+    /* Frame is fully decoded. Input exhausted and output sitting in buffer. */
+    if (0 == zresult) {
+        self->finishedInput = 1;
+        self->finishedOutput = 1;
+    }
+
+    /* If it produced output data, return it. */
+    if (self->output.pos) {
+        if (self->output.pos < self->outSize) {
+            if (safe_pybytes_resize(&chunk, self->output.pos)) {
+                Py_XDECREF(chunk);
+                result.errored = 1;
+                return result;
+            }
+        }
+    }
+    else {
+        Py_DECREF(chunk);
+        chunk = NULL;
+    }
+
+    result.errored = 0;
+    result.chunk = chunk;
+
+    return result;
+}
+
+static PyObject *
+ZstdDecompressorIterator_iternext(ZstdDecompressorIterator *self) {
+    PyObject *readResult = NULL;
+    char *readBuffer;
+    Py_ssize_t readSize;
+    Py_ssize_t bufferRemaining;
+    DecompressorIteratorResult result;
+
+    if (self->finishedOutput) {
+        PyErr_SetString(PyExc_StopIteration, "output flushed");
+        return NULL;
+    }
+
+    /* If we have data left in the input, consume it. */
+    if (self->input.pos < self->input.size) {
+        result = read_decompressor_iterator(self);
+        if (result.chunk || result.errored) {
+            return result.chunk;
+        }
+
+        /* Else fall through to get more data from input. */
+    }
+
+read_from_source:
+
+    if (!self->finishedInput) {
+        if (self->reader) {
+            readResult =
+                PyObject_CallMethod(self->reader, "read", "I", self->inSize);
+            if (!readResult) {
+                return NULL;
+            }
+
+            PyBytes_AsStringAndSize(readResult, &readBuffer, &readSize);
+        }
+        else {
+            assert(self->buffer.buf);
+
+            /* Only support contiguous C arrays for now */
+            assert(self->buffer.strides == NULL &&
+                   self->buffer.suboffsets == NULL);
+            assert(self->buffer.itemsize == 1);
+
+            /* TODO avoid memcpy() below */
+            readBuffer = (char *)self->buffer.buf + self->bufferOffset;
+            bufferRemaining = self->buffer.len - self->bufferOffset;
+            readSize = min(bufferRemaining, (Py_ssize_t)self->inSize);
+            self->bufferOffset += readSize;
+        }
+
+        if (readSize) {
+            if (!self->readCount && self->skipBytes) {
+                assert(self->skipBytes < self->inSize);
+                if ((Py_ssize_t)self->skipBytes >= readSize) {
+                    PyErr_SetString(PyExc_ValueError,
+                                    "skip_bytes larger than first input chunk; "
+                                    "this scenario is currently unsupported");
+                    Py_XDECREF(readResult);
+                    return NULL;
+                }
+
+                readBuffer = readBuffer + self->skipBytes;
+                readSize -= self->skipBytes;
+            }
+
+            /* Copy input into previously allocated buffer because it can live
+            longer than a single function call and we don't want to keep a ref
+            to a Python object around. This could be changed... */
+            memcpy((void *)self->input.src, readBuffer, readSize);
+            self->input.size = readSize;
+            self->input.pos = 0;
+        }
+        /* No bytes on first read must mean an empty input stream. */
+        else if (!self->readCount) {
+            self->finishedInput = 1;
+            self->finishedOutput = 1;
+            Py_XDECREF(readResult);
+            PyErr_SetString(PyExc_StopIteration, "empty input");
+            return NULL;
+        }
+        else {
+            self->finishedInput = 1;
+        }
+
+        /* We've copied the data managed by memory. Discard the Python object.
+         */
+        Py_XDECREF(readResult);
+    }
+
+    result = read_decompressor_iterator(self);
+    if (result.errored || result.chunk) {
+        return result.chunk;
+    }
+
+    /* No new output data. Try again unless we know there is no more data. */
+    if (!self->finishedInput) {
+        goto read_from_source;
+    }
+
+    PyErr_SetString(PyExc_StopIteration, "input exhausted");
+    return NULL;
+}
+
+PyType_Slot ZstdDecompressorIteratorSlots[] = {
+    {Py_tp_dealloc, ZstdDecompressorIterator_dealloc},
+    {Py_tp_iter, ZstdDecompressorIterator_iter},
+    {Py_tp_iternext, ZstdDecompressorIterator_iternext},
+    {Py_tp_new, PyType_GenericNew},
+    {0, NULL},
+};
+
+PyType_Spec ZstdDecompressorIteratorSpec = {
+    "zstd.ZstdDecompressorIterator",
+    sizeof(ZstdDecompressorIterator),
+    0,
+    Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,
+    ZstdDecompressorIteratorSlots,
+};
+
+PyTypeObject *ZstdDecompressorIteratorType;
+
+void decompressoriterator_module_init(PyObject *mod) {
+    ZstdDecompressorIteratorType =
+        (PyTypeObject *)PyType_FromSpec(&ZstdDecompressorIteratorSpec);
+    if (PyType_Ready(ZstdDecompressorIteratorType) < 0) {
+        return;
+    }
+}
diff --git a/contrib/python/zstandard/py3/c-ext/frameparams.c b/contrib/python/zstandard/py3/c-ext/frameparams.c
new file mode 100644
index 00000000000..951a91d6319
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/frameparams.c
@@ -0,0 +1,97 @@
+/**
+ * Copyright (c) 2017-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#include "python-zstandard.h"
+
+extern PyObject *ZstdError;
+
+FrameParametersObject *get_frame_parameters(PyObject *self, PyObject *args,
+                                            PyObject *kwargs) {
+    static char *kwlist[] = {"data", NULL};
+
+    Py_buffer source;
+    ZSTD_frameHeader header;
+    FrameParametersObject *result = NULL;
+    size_t zresult;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*:get_frame_parameters",
+                                     kwlist, &source)) {
+        return NULL;
+    }
+
+    zresult = ZSTD_getFrameHeader(&header, source.buf, source.len);
+
+    if (ZSTD_isError(zresult)) {
+        PyErr_Format(ZstdError, "cannot get frame parameters: %s",
+                     ZSTD_getErrorName(zresult));
+        goto finally;
+    }
+
+    if (zresult) {
+        PyErr_Format(ZstdError,
+                     "not enough data for frame parameters; need %zu bytes",
+                     zresult);
+        goto finally;
+    }
+
+    result = PyObject_New(FrameParametersObject, FrameParametersType);
+    if (!result) {
+        goto finally;
+    }
+
+    result->frameContentSize = header.frameContentSize;
+    result->windowSize = header.windowSize;
+    result->dictID = header.dictID;
+    result->checksumFlag = header.checksumFlag ? 1 : 0;
+
+finally:
+    PyBuffer_Release(&source);
+    return result;
+}
+
+static void FrameParameters_dealloc(PyObject *self) {
+    PyObject_Del(self);
+}
+
+static PyMemberDef FrameParameters_members[] = {
+    {"content_size", T_ULONGLONG,
+     offsetof(FrameParametersObject, frameContentSize), READONLY,
+     "frame content size"},
+    {"window_size", T_ULONGLONG, offsetof(FrameParametersObject, windowSize),
+     READONLY, "window size"},
+    {"dict_id", T_UINT, offsetof(FrameParametersObject, dictID), READONLY,
+     "dictionary ID"},
+    {"has_checksum", T_BOOL, offsetof(FrameParametersObject, checksumFlag),
+     READONLY, "checksum flag"},
+    {NULL}};
+
+PyType_Slot FrameParametersSlots[] = {
+    {Py_tp_dealloc, FrameParameters_dealloc},
+    {Py_tp_members, FrameParameters_members},
+    {0, NULL},
+};
+
+PyType_Spec FrameParametersSpec = {
+    "zstd.FrameParameters",
+    sizeof(FrameParametersObject),
+    0,
+    Py_TPFLAGS_DEFAULT,
+    FrameParametersSlots,
+};
+
+PyTypeObject *FrameParametersType;
+
+void frameparams_module_init(PyObject *mod) {
+    FrameParametersType = (PyTypeObject *)PyType_FromSpec(&FrameParametersSpec);
+    if (PyType_Ready(FrameParametersType) < 0) {
+        return;
+    }
+
+    Py_INCREF(FrameParametersType);
+    PyModule_AddObject(mod, "FrameParameters", (PyObject *)FrameParametersType);
+}
diff --git a/contrib/python/zstandard/py3/c-ext/python-zstandard.h b/contrib/python/zstandard/py3/c-ext/python-zstandard.h
new file mode 100644
index 00000000000..a58261d0e40
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/python-zstandard.h
@@ -0,0 +1,392 @@
+/**
+ * Copyright (c) 2016-present, Gregory Szorc
+ * All rights reserved.
+ *
+ * This software may be modified and distributed under the terms
+ * of the BSD license. See the LICENSE file for details.
+ */
+
+#ifndef PYTHON_ZSTANDARD_H
+#define PYTHON_ZSTANDARD_H
+
+#define PY_SSIZE_T_CLEAN
+#include <Python.h>
+#include <pythoncapi_compat.h>
+#include "structmember.h"
+
+#define ZSTD_STATIC_LINKING_ONLY
+#define ZDICT_STATIC_LINKING_ONLY
+
+#ifdef ZSTD_SINGLE_FILE
+#error #include <zstd.c>
+
+/* We use private APIs from pool.h. We can't rely on availability
+   of this header or symbols when linking against the system libzstd.
+   But we know it works when using the bundled single file library. */
+#define HAVE_ZSTD_POOL_APIS
+#else
+#include <zdict.h>
+#include <zstd.h>
+#endif
+
+/* Remember to change the string in zstandard/__init__.py, rust-ext/src/lib.rs,
+   and debian/changelog as well */
+#define PYTHON_ZSTANDARD_VERSION "0.21.0"
+
+typedef enum {
+    compressorobj_flush_finish,
+    compressorobj_flush_block,
+} CompressorObj_Flush;
+
+/*
+   Represents a ZstdCompressionParameters type.
+
+   This type holds all the low-level compression parameters that can be set.
+*/
+typedef struct {
+    PyObject_HEAD ZSTD_CCtx_params *params;
+} ZstdCompressionParametersObject;
+
+extern PyTypeObject *ZstdCompressionParametersType;
+
+/*
+   Represents a FrameParameters type.
+
+   This type is basically a wrapper around ZSTD_frameParams.
+*/
+typedef struct {
+    PyObject_HEAD unsigned long long frameContentSize;
+    unsigned long long windowSize;
+    unsigned dictID;
+    char checksumFlag;
+} FrameParametersObject;
+
+extern PyTypeObject *FrameParametersType;
+
+/*
+   Represents a ZstdCompressionDict type.
+
+   Instances hold data used for a zstd compression dictionary.
+*/
+typedef struct {
+    PyObject_HEAD
+
+        /* Pointer to dictionary data. Owned by self. */
+        void *dictData;
+    /* Size of dictionary data. */
+    size_t dictSize;
+    ZSTD_dictContentType_e dictType;
+    /* k parameter for cover dictionaries. Only populated by train_cover_dict().
+     */
+    unsigned k;
+    /* d parameter for cover dictionaries. Only populated by train_cover_dict().
+     */
+    unsigned d;
+    /* Digested dictionary, suitable for reuse. */
+    ZSTD_CDict *cdict;
+    ZSTD_DDict *ddict;
+} ZstdCompressionDict;
+
+extern PyTypeObject *ZstdCompressionDictType;
+
+/*
+   Represents a ZstdCompressor type.
+*/
+typedef struct {
+    PyObject_HEAD
+
+        /* Number of threads to use for operations. */
+        unsigned int threads;
+    /* Pointer to compression dictionary to use. NULL if not using dictionary
+       compression. */
+    ZstdCompressionDict *dict;
+    /* Compression context to use. Populated during object construction. */
+    ZSTD_CCtx *cctx;
+    /* Compression parameters in use. */
+    ZSTD_CCtx_params *params;
+} ZstdCompressor;
+
+extern PyTypeObject *ZstdCompressorType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdCompressor *compressor;
+    ZSTD_outBuffer output;
+    int finished;
+} ZstdCompressionObj;
+
+extern PyTypeObject *ZstdCompressionObjType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdCompressor *compressor;
+    PyObject *writer;
+    ZSTD_outBuffer output;
+    size_t outSize;
+    int entered;
+    int closing;
+    char closed;
+    int writeReturnRead;
+    int closefd;
+    unsigned long long bytesCompressed;
+} ZstdCompressionWriter;
+
+extern PyTypeObject *ZstdCompressionWriterType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdCompressor *compressor;
+    PyObject *reader;
+    Py_buffer buffer;
+    Py_ssize_t bufferOffset;
+    size_t inSize;
+    size_t outSize;
+
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    int finishedOutput;
+    int finishedInput;
+    PyObject *readResult;
+} ZstdCompressorIterator;
+
+extern PyTypeObject *ZstdCompressorIteratorType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdCompressor *compressor;
+    PyObject *reader;
+    Py_buffer buffer;
+    size_t readSize;
+    int closefd;
+
+    int entered;
+    char closed;
+    unsigned long long bytesCompressed;
+
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    int finishedInput;
+    int finishedOutput;
+    PyObject *readResult;
+} ZstdCompressionReader;
+
+extern PyTypeObject *ZstdCompressionReaderType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdCompressor *compressor;
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    Py_buffer inBuffer;
+    int finished;
+    size_t chunkSize;
+} ZstdCompressionChunker;
+
+extern PyTypeObject *ZstdCompressionChunkerType;
+
+typedef enum {
+    compressionchunker_mode_normal,
+    compressionchunker_mode_flush,
+    compressionchunker_mode_finish,
+} CompressionChunkerMode;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdCompressionChunker *chunker;
+    CompressionChunkerMode mode;
+} ZstdCompressionChunkerIterator;
+
+extern PyTypeObject *ZstdCompressionChunkerIteratorType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZSTD_DCtx *dctx;
+    ZstdCompressionDict *dict;
+    size_t maxWindowSize;
+    ZSTD_format_e format;
+} ZstdDecompressor;
+
+extern PyTypeObject *ZstdDecompressorType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdDecompressor *decompressor;
+    size_t outSize;
+    int finished;
+    PyObject *unused_data;
+} ZstdDecompressionObj;
+
+extern PyTypeObject *ZstdDecompressionObjType;
+
+typedef struct {
+    PyObject_HEAD
+
+        /* Parent decompressor to which this object is associated. */
+        ZstdDecompressor *decompressor;
+    /* Object to read() from (if reading from a stream). */
+    PyObject *reader;
+    /* Size for read() operations on reader. */
+    size_t readSize;
+    /* Whether a read() can return data spanning multiple zstd frames. */
+    int readAcrossFrames;
+    /* Buffer to read from (if reading from a buffer). */
+    Py_buffer buffer;
+    /* Whether to close the inner object on close() */
+    int closefd;
+
+    /* Whether the context manager is active. */
+    int entered;
+    /* Whether we've closed the stream. */
+    char closed;
+
+    /* Number of bytes decompressed and returned to user. */
+    unsigned long long bytesDecompressed;
+
+    /* Tracks data going into decompressor. */
+    ZSTD_inBuffer input;
+
+    /* Holds output from read() operation on reader. */
+    PyObject *readResult;
+
+    /* Whether all input has been sent to the decompressor. */
+    int finishedInput;
+    /* Whether all output has been flushed from the decompressor. */
+    int finishedOutput;
+} ZstdDecompressionReader;
+
+extern PyTypeObject *ZstdDecompressionReaderType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdDecompressor *decompressor;
+    PyObject *writer;
+    size_t outSize;
+    int entered;
+    int closing;
+    char closed;
+    int writeReturnRead;
+    int closefd;
+} ZstdDecompressionWriter;
+
+extern PyTypeObject *ZstdDecompressionWriterType;
+
+typedef struct {
+    PyObject_HEAD
+
+        ZstdDecompressor *decompressor;
+    PyObject *reader;
+    Py_buffer buffer;
+    Py_ssize_t bufferOffset;
+    size_t inSize;
+    size_t outSize;
+    size_t skipBytes;
+    ZSTD_inBuffer input;
+    ZSTD_outBuffer output;
+    Py_ssize_t readCount;
+    int finishedInput;
+    int finishedOutput;
+} ZstdDecompressorIterator;
+
+extern PyTypeObject *ZstdDecompressorIteratorType;
+
+typedef struct {
+    int errored;
+    PyObject *chunk;
+} DecompressorIteratorResult;
+
+typedef struct {
+    /* The public API is that these are 64-bit unsigned integers. So these can't
+     * be size_t, even though values larger than SIZE_MAX or PY_SSIZE_T_MAX may
+     * be nonsensical for this platform. */
+    unsigned long long offset;
+    unsigned long long length;
+} BufferSegment;
+
+typedef struct {
+    PyObject_HEAD
+
+        PyObject *parent;
+    BufferSegment *segments;
+    Py_ssize_t segmentCount;
+} ZstdBufferSegments;
+
+extern PyTypeObject *ZstdBufferSegmentsType;
+
+typedef struct {
+    PyObject_HEAD
+
+        PyObject *parent;
+    void *data;
+    Py_ssize_t dataSize;
+    unsigned long long offset;
+} ZstdBufferSegment;
+
+extern PyTypeObject *ZstdBufferSegmentType;
+
+typedef struct {
+    PyObject_HEAD
+
+        Py_buffer parent;
+    void *data;
+    unsigned long long dataSize;
+    BufferSegment *segments;
+    Py_ssize_t segmentCount;
+    int useFree;
+} ZstdBufferWithSegments;
+
+extern PyTypeObject *ZstdBufferWithSegmentsType;
+
+/**
+ * An ordered collection of BufferWithSegments exposed as a squashed collection.
+ *
+ * This type provides a virtual view spanning multiple BufferWithSegments
+ * instances. It allows multiple instances to be "chained" together and
+ * exposed as a single collection. e.g. if there are 2 buffers holding
+ * 10 segments each, then o[14] will access the 5th segment in the 2nd buffer.
+ */
+typedef struct {
+    PyObject_HEAD
+
+        /* An array of buffers that should be exposed through this instance. */
+        ZstdBufferWithSegments **buffers;
+    /* Number of elements in buffers array. */
+    Py_ssize_t bufferCount;
+    /* Array of first offset in each buffer instance. 0th entry corresponds
+       to number of elements in the 0th buffer. 1st entry corresponds to the
+       sum of elements in 0th and 1st buffers. */
+    Py_ssize_t *firstElements;
+} ZstdBufferWithSegmentsCollection;
+
+extern PyTypeObject *ZstdBufferWithSegmentsCollectionType;
+
+int set_parameter(ZSTD_CCtx_params *params, ZSTD_cParameter param, int value);
+int set_parameters(ZSTD_CCtx_params *params,
+                   ZstdCompressionParametersObject *obj);
+int to_cparams(ZstdCompressionParametersObject *params,
+               ZSTD_compressionParameters *cparams);
+FrameParametersObject *get_frame_parameters(PyObject *self, PyObject *args,
+                                            PyObject *kwargs);
+int ensure_ddict(ZstdCompressionDict *dict);
+int ensure_dctx(ZstdDecompressor *decompressor, int loadDict);
+ZstdCompressionDict *train_dictionary(PyObject *self, PyObject *args,
+                                      PyObject *kwargs);
+ZstdBufferWithSegments *
+BufferWithSegments_FromMemory(void *data, unsigned long long dataSize,
+                              BufferSegment *segments, Py_ssize_t segmentsSize);
+Py_ssize_t
+BufferWithSegmentsCollection_length(ZstdBufferWithSegmentsCollection *);
+int cpu_count(void);
+size_t roundpow2(size_t);
+int safe_pybytes_resize(PyObject **obj, Py_ssize_t size);
+void set_io_unsupported_operation(void);
+
+#endif
diff --git a/contrib/python/zstandard/py3/c-ext/pythoncapi_compat.h b/contrib/python/zstandard/py3/c-ext/pythoncapi_compat.h
new file mode 100644
index 00000000000..450d7ed6e80
--- /dev/null
+++ b/contrib/python/zstandard/py3/c-ext/pythoncapi_compat.h
@@ -0,0 +1,307 @@
+// Header file providing new functions of the Python C API to old Python
+// versions.
+//
+// File distributed under the MIT license.
+//
+// Homepage:
+// https://github.com/pythoncapi/pythoncapi_compat
+//
+// Latest version:
+// https://raw.githubusercontent.com/pythoncapi/pythoncapi_compat/master/pythoncapi_compat.h
+//
+// SPDX-License-Identifier: MIT
+
+#ifndef PYTHONCAPI_COMPAT
+#define PYTHONCAPI_COMPAT
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <Python.h>
+#include "frameobject.h"          // PyFrameObject, PyFrame_GetBack()
+
+
+// Compatibility with Visual Studio 2013 and older which don't support
+// the inline keyword in C (only in C++): use __inline instead.
+#if (defined(_MSC_VER) && _MSC_VER < 1900 \
+     && !defined(__cplusplus) && !defined(inline))
+#  define inline __inline
+#  define PYTHONCAPI_COMPAT_MSC_INLINE
+   // These two macros are undefined at the end of this file
+#endif
+
+
+// Cast argument to PyObject* type.
+#ifndef _PyObject_CAST
+#  define _PyObject_CAST(op) ((PyObject*)(op))
+#endif
+#ifndef _PyObject_CAST_CONST
+#  define _PyObject_CAST_CONST(op) ((const PyObject*)(op))
+#endif
+
+
+// bpo-42262 added Py_NewRef() to Python 3.10.0a3
+#if PY_VERSION_HEX < 0x030A00A3 && !defined(Py_NewRef)
+static inline PyObject* _Py_NewRef(PyObject *obj)
+{
+    Py_INCREF(obj);
+    return obj;
+}
+#define Py_NewRef(obj) _Py_NewRef(_PyObject_CAST(obj))
+#endif
+
+
+// bpo-42262 added Py_XNewRef() to Python 3.10.0a3
+#if PY_VERSION_HEX < 0x030A00A3 && !defined(Py_XNewRef)
+static inline PyObject* _Py_XNewRef(PyObject *obj)
+{
+    Py_XINCREF(obj);
+    return obj;
+}
+#define Py_XNewRef(obj) _Py_XNewRef(_PyObject_CAST(obj))
+#endif
+
+
+// bpo-39573 added Py_SET_REFCNT() to Python 3.9.0a4
+#if PY_VERSION_HEX < 0x030900A4 && !defined(Py_SET_REFCNT)
+static inline void _Py_SET_REFCNT(PyObject *ob, Py_ssize_t refcnt)
+{
+    ob->ob_refcnt = refcnt;
+}
+#define Py_SET_REFCNT(ob, refcnt) _Py_SET_REFCNT(_PyObject_CAST(ob), refcnt)
+#endif
+
+
+// bpo-39573 added Py_SET_TYPE() to Python 3.9.0a4
+#if PY_VERSION_HEX < 0x030900A4 && !defined(Py_SET_TYPE)
+static inline void
+_Py_SET_TYPE(PyObject *ob, PyTypeObject *type)
+{
+    ob->ob_type = type;
+}
+#define Py_SET_TYPE(ob, type) _Py_SET_TYPE(_PyObject_CAST(ob), type)
+#endif
+
+
+// bpo-39573 added Py_SET_SIZE() to Python 3.9.0a4
+#if PY_VERSION_HEX < 0x030900A4 && !defined(Py_SET_SIZE)
+static inline void
+_Py_SET_SIZE(PyVarObject *ob, Py_ssize_t size)
+{
+    ob->ob_size = size;
+}
+#define Py_SET_SIZE(ob, size) _Py_SET_SIZE((PyVarObject*)(ob), size)
+#endif
+
+
+// bpo-40421 added PyFrame_GetCode() to Python 3.9.0b1
+#if PY_VERSION_HEX < 0x030900B1
+static inline PyCodeObject*
+PyFrame_GetCode(PyFrameObject *frame)
+{
+    PyCodeObject *code;
+    assert(frame != NULL);
+    code = frame->f_code;
+    assert(code != NULL);
+    Py_INCREF(code);
+    return code;
+}
+#endif
+
+static inline PyCodeObject*
+_PyFrame_GetCodeBorrow(PyFrameObject *frame)
+{
+    PyCodeObject *code = PyFrame_GetCode(frame);
+    Py_DECREF(code);
+    return code;  // borrowed reference
+}
+
+
+// bpo-40421 added PyFrame_GetCode() to Python 3.9.0b1
+#if PY_VERSION_HEX < 0x030900B1
+static inline PyFrameObject*
+PyFrame_GetBack(PyFrameObject *frame)
+{
+    PyFrameObject *back;
+    assert(frame != NULL);
+    back = frame->f_back;
+    Py_XINCREF(back);
+    return back;
+}
+#endif
+
+static inline PyFrameObject*
+_PyFrame_GetBackBorrow(PyFrameObject *frame)
+{
+    PyFrameObject *back = PyFrame_GetBack(frame);
+    Py_XDECREF(back);
+    return back;  // borrowed reference
+}
+
+
+// bpo-39947 added PyThreadState_GetInterpreter() to Python 3.9.0a5
+#if PY_VERSION_HEX < 0x030900A5
+static inline PyInterpreterState *
+PyThreadState_GetInterpreter(PyThreadState *tstate)
+{
+    assert(tstate != NULL);
+    return tstate->interp;
+}
+#endif
+
+
+// bpo-40429 added PyThreadState_GetFrame() to Python 3.9.0b1
+#if PY_VERSION_HEX < 0x030900B1
+static inline PyFrameObject*
+PyThreadState_GetFrame(PyThreadState *tstate)
+{
+    PyFrameObject *frame;
+    assert(tstate != NULL);
+    frame = tstate->frame;
+    Py_XINCREF(frame);
+    return frame;
+}
+#endif
+
+static inline PyFrameObject*
+_PyThreadState_GetFrameBorrow(PyThreadState *tstate)
+{
+    PyFrameObject *frame = PyThreadState_GetFrame(tstate);
+    Py_XDECREF(frame);
+    return frame;  // borrowed reference
+}
+
+
+// bpo-39947 added PyInterpreterState_Get() to Python 3.9.0a5
+#if PY_VERSION_HEX < 0x030900A5
+static inline PyInterpreterState *
+PyInterpreterState_Get(void)
+{
+    PyThreadState *tstate;
+    PyInterpreterState *interp;
+
+    tstate = PyThreadState_GET();
+    if (tstate == NULL) {
+        Py_FatalError("GIL released (tstate is NULL)");
+    }
+    interp = tstate->interp;
+    if (interp == NULL) {
+        Py_FatalError("no current interpreter");
+    }
+    return interp;
+}
+#endif
+
+
+// bpo-39947 added PyInterpreterState_Get() to Python 3.9.0a6
+#if 0x030700A1 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x030900A6
+static inline uint64_t
+PyThreadState_GetID(PyThreadState *tstate)
+{
+    assert(tstate != NULL);
+    return tstate->id;
+}
+#endif
+
+
+// bpo-37194 added PyObject_CallNoArgs() to Python 3.9.0a1
+#if PY_VERSION_HEX < 0x030900A1
+static inline PyObject*
+PyObject_CallNoArgs(PyObject *func)
+{
+    return PyObject_CallFunctionObjArgs(func, NULL);
+}
+#endif
+
+
+// bpo-39245 made PyObject_CallOneArg() public (previously called
+// _PyObject_CallOneArg) in Python 3.9.0a4
+#if PY_VERSION_HEX < 0x030900A4
+static inline PyObject*
+PyObject_CallOneArg(PyObject *func, PyObject *arg)
+{
+    return PyObject_CallFunctionObjArgs(func, arg, NULL);
+}
+#endif
+
+
+// bpo-1635741 added PyModule_AddObjectRef() to Python 3.10.0a3
+#if PY_VERSION_HEX < 0x030A00A3
+static inline int
+PyModule_AddObjectRef(PyObject *module, const char *name, PyObject *value)
+{
+    Py_XINCREF(value);
+    int res = PyModule_AddObject(module, name, value);
+    if (res < 0) {
+        Py_XDECREF(value);
+    }
+    return res;
+}
+#endif
+
+
+// bpo-40024 added PyModule_AddType() to Python 3.9.0a5
+#if PY_VERSION_HEX < 0x030900A5
+static inline int
+PyModule_AddType(PyObject *module, PyTypeObject *type)
+{
+    const char *name, *dot;
+
+    if (PyType_Ready(type) < 0) {
+        return -1;
+    }
+
+    // inline _PyType_Name()
+    name = type->tp_name;
+    assert(name != NULL);
+    dot = strrchr(name, '.');
+    if (dot != NULL) {
+        name = dot + 1;
+    }
+
+    return PyModule_AddObjectRef(module, name, (PyObject *)type);
+}
+#endif
+
+
+// bpo-40241 added PyObject_GC_IsTracked() to Python 3.9.0a6.
+// bpo-4688 added _PyObject_GC_IS_TRACKED() to Python 2.7.0a2.
+#if PY_VERSION_HEX < 0x030900A6
+static inline int
+PyObject_GC_IsTracked(PyObject* obj)
+{
+    return (PyObject_IS_GC(obj) && _PyObject_GC_IS_TRACKED(obj));
+}
+#endif
+
+// bpo-40241 added PyObject_GC_IsFinalized() to Python 3.9.0a6.
+// bpo-18112 added _PyGCHead_FINALIZED() to Python 3.4.0 final.
+#if PY_VERSION_HEX < 0x030900A6 && PY_VERSION_HEX >= 0x030400F0
+static inline int
+PyObject_GC_IsFinalized(PyObject *obj)
+{
+    return (PyObject_IS_GC(obj) && _PyGCHead_FINALIZED((PyGC_Head *)(obj)-1));
+}
+#endif
+
+
+// bpo-39573 added Py_IS_TYPE() to Python 3.9.0a4
+#if PY_VERSION_HEX < 0x030900A4 && !defined(Py_IS_TYPE)
+static inline int
+_Py_IS_TYPE(const PyObject *ob, const PyTypeObject *type) {
+    return ob->ob_type == type;
+}
+#define Py_IS_TYPE(ob, type) _Py_IS_TYPE(_PyObject_CAST_CONST(ob), type)
+#endif
+
+
+#ifdef PYTHONCAPI_COMPAT_MSC_INLINE
+#  undef inline
+#  undef PYTHONCAPI_COMPAT_MSC_INLINE
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+#endif  // PYTHONCAPI_COMPAT
diff --git a/contrib/python/zstandard/py3/ya.make b/contrib/python/zstandard/py3/ya.make
new file mode 100644
index 00000000000..27e56f1470d
--- /dev/null
+++ b/contrib/python/zstandard/py3/ya.make
@@ -0,0 +1,59 @@
+# Generated by devtools/yamaker (pypi).
+
+PY3_LIBRARY()
+
+VERSION(0.21.0)
+
+LICENSE(BSD-3-Clause)
+
+PEERDIR(
+    contrib/libs/zstd
+)
+
+ADDINCL(
+    contrib/libs/zstd/include
+    contrib/python/zstandard/py3/c-ext
+)
+
+NO_COMPILER_WARNINGS()
+
+NO_LINT()
+
+SRCS(
+    c-ext/backend_c.c
+    c-ext/bufferutil.c
+    c-ext/compressionchunker.c
+    c-ext/compressiondict.c
+    c-ext/compressionparams.c
+    c-ext/compressionreader.c
+    c-ext/compressionwriter.c
+    c-ext/compressobj.c
+    c-ext/compressor.c
+    c-ext/compressoriterator.c
+    c-ext/constants.c
+    c-ext/decompressionreader.c
+    c-ext/decompressionwriter.c
+    c-ext/decompressobj.c
+    c-ext/decompressor.c
+    c-ext/decompressoriterator.c
+    c-ext/frameparams.c
+)
+
+PY_REGISTER(
+    zstandard.backend_c
+)
+
+PY_SRCS(
+    TOP_LEVEL
+    zstandard/__init__.py
+    zstandard/__init__.pyi
+)
+
+RESOURCE_FILES(
+    PREFIX contrib/python/zstandard/py3/
+    .dist-info/METADATA
+    .dist-info/top_level.txt
+    zstandard/py.typed
+)
+
+END()
diff --git a/contrib/python/zstandard/py3/zstandard/__init__.py b/contrib/python/zstandard/py3/zstandard/__init__.py
new file mode 100644
index 00000000000..7840473e1cb
--- /dev/null
+++ b/contrib/python/zstandard/py3/zstandard/__init__.py
@@ -0,0 +1,210 @@
+# Copyright (c) 2017-present, Gregory Szorc
+# All rights reserved.
+#
+# This software may be modified and distributed under the terms
+# of the BSD license. See the LICENSE file for details.
+
+"""Python interface to the Zstandard (zstd) compression library."""
+
+from __future__ import absolute_import, unicode_literals
+
+# This module serves 2 roles:
+#
+# 1) Export the C or CFFI "backend" through a central module.
+# 2) Implement additional functionality built on top of C or CFFI backend.
+
+import builtins
+import io
+import os
+import platform
+
+from typing import ByteString
+
+# Some Python implementations don't support C extensions. That's why we have
+# a CFFI implementation in the first place. The code here import one of our
+# "backends" then re-exports the symbols from this module. For convenience,
+# we support falling back to the CFFI backend if the C extension can't be
+# imported. But for performance reasons, we only do this on unknown Python
+# implementation. Notably, for CPython we require the C extension by default.
+# Because someone will inevitably want special behavior, the behavior is
+# configurable via an environment variable. A potentially better way to handle
+# this is to import a special ``__importpolicy__`` module or something
+# defining a variable and `setup.py` could write the file with whatever
+# policy was specified at build time. Until someone needs it, we go with
+# the hacky but simple environment variable approach.
+_module_policy = os.environ.get("PYTHON_ZSTANDARD_IMPORT_POLICY", "default")
+
+if _module_policy == "default":
+    if platform.python_implementation() in ("CPython",):
+        from .backend_c import *  # type: ignore
+
+        backend = "cext"
+    elif platform.python_implementation() in ("PyPy",):
+        from .backend_cffi import *  # type: ignore
+
+        backend = "cffi"
+    else:
+        try:
+            from .backend_c import *
+
+            backend = "cext"
+        except ImportError:
+            from .backend_cffi import *
+
+            backend = "cffi"
+elif _module_policy == "cffi_fallback":
+    try:
+        from .backend_c import *
+
+        backend = "cext"
+    except ImportError:
+        from .backend_cffi import *
+
+        backend = "cffi"
+elif _module_policy == "rust":
+    from .backend_rust import *  # type: ignore
+
+    backend = "rust"
+elif _module_policy == "cext":
+    from .backend_c import *
+
+    backend = "cext"
+elif _module_policy == "cffi":
+    from .backend_cffi import *
+
+    backend = "cffi"
+else:
+    raise ImportError(
+        "unknown module import policy: %s; use default, cffi_fallback, "
+        "cext, or cffi" % _module_policy
+    )
+
+# Keep this in sync with python-zstandard.h, rust-ext/src/lib.rs, and debian/changelog.
+__version__ = "0.21.0"
+
+_MODE_CLOSED = 0
+_MODE_READ = 1
+_MODE_WRITE = 2
+
+
+def open(
+    filename,
+    mode="rb",
+    cctx=None,
+    dctx=None,
+    encoding=None,
+    errors=None,
+    newline=None,
+    closefd=None,
+):
+    """Create a file object with zstd (de)compression.
+
+    The object returned from this function will be a
+    :py:class:`ZstdDecompressionReader` if opened for reading in binary mode,
+    a :py:class:`ZstdCompressionWriter` if opened for writing in binary mode,
+    or an ``io.TextIOWrapper`` if opened for reading or writing in text mode.
+
+    :param filename:
+       ``bytes``, ``str``, or ``os.PathLike`` defining a file to open or a
+       file object (with a ``read()`` or ``write()`` method).
+    :param mode:
+       ``str`` File open mode. Accepts any of the open modes recognized by
+       ``open()``.
+    :param cctx:
+       ``ZstdCompressor`` to use for compression. If not specified and file
+       is opened for writing, the default ``ZstdCompressor`` will be used.
+    :param dctx:
+       ``ZstdDecompressor`` to use for decompression. If not specified and file
+       is opened for reading, the default ``ZstdDecompressor`` will be used.
+    :param encoding:
+        ``str`` that defines text encoding to use when file is opened in text
+        mode.
+    :param errors:
+       ``str`` defining text encoding error handling mode.
+    :param newline:
+       ``str`` defining newline to use in text mode.
+    :param closefd:
+       ``bool`` whether to close the file when the returned object is closed.
+        Only used if a file object is passed. If a filename is specified, the
+        opened file is always closed when the returned object is closed.
+    """
+    normalized_mode = mode.replace("t", "")
+
+    if normalized_mode in ("r", "rb"):
+        dctx = dctx or ZstdDecompressor()
+        open_mode = "r"
+        raw_open_mode = "rb"
+    elif normalized_mode in ("w", "wb", "a", "ab", "x", "xb"):
+        cctx = cctx or ZstdCompressor()
+        open_mode = "w"
+        raw_open_mode = normalized_mode
+        if not raw_open_mode.endswith("b"):
+            raw_open_mode = raw_open_mode + "b"
+    else:
+        raise ValueError("Invalid mode: {!r}".format(mode))
+
+    if hasattr(os, "PathLike"):
+        types = (str, bytes, os.PathLike)
+    else:
+        types = (str, bytes)
+
+    if isinstance(filename, types):  # type: ignore
+        inner_fh = builtins.open(filename, raw_open_mode)
+        closefd = True
+    elif hasattr(filename, "read") or hasattr(filename, "write"):
+        inner_fh = filename
+        closefd = bool(closefd)
+    else:
+        raise TypeError(
+            "filename must be a str, bytes, file or PathLike object"
+        )
+
+    if open_mode == "r":
+        fh = dctx.stream_reader(inner_fh, closefd=closefd)
+    elif open_mode == "w":
+        fh = cctx.stream_writer(inner_fh, closefd=closefd)
+    else:
+        raise RuntimeError("logic error in zstandard.open() handling open mode")
+
+    if "b" not in normalized_mode:
+        return io.TextIOWrapper(
+            fh, encoding=encoding, errors=errors, newline=newline
+        )
+    else:
+        return fh
+
+
+def compress(data: ByteString, level: int = 3) -> bytes:
+    """Compress source data using the zstd compression format.
+
+    This performs one-shot compression using basic/default compression
+    settings.
+
+    This method is provided for convenience and is equivalent to calling
+    ``ZstdCompressor(level=level).compress(data)``.
+
+    If you find yourself calling this function in a tight loop,
+    performance will be greater if you construct a single ``ZstdCompressor``
+    and repeatedly call ``compress()`` on it.
+    """
+    cctx = ZstdCompressor(level=level)
+
+    return cctx.compress(data)
+
+
+def decompress(data: ByteString, max_output_size: int = 0) -> bytes:
+    """Decompress a zstd frame into its original data.
+
+    This performs one-shot decompression using basic/default compression
+    settings.
+
+    This method is provided for convenience and is equivalent to calling
+    ``ZstdDecompressor().decompress(data, max_output_size=max_output_size)``.
+
+    If you find yourself calling this function in a tight loop, performance
+    will be greater if you construct a single ``ZstdDecompressor`` and
+    repeatedly call ``decompress()`` on it.
+    """
+    dctx = ZstdDecompressor()
+
+    return dctx.decompress(data, max_output_size=max_output_size)
diff --git a/contrib/python/zstandard/py3/zstandard/py.typed b/contrib/python/zstandard/py3/zstandard/py.typed
new file mode 100644
index 00000000000..e69de29bb2d
--- /dev/null
+++ b/contrib/python/zstandard/py3/zstandard/py.typed
diff --git a/contrib/python/zstandard/ya.make b/contrib/python/zstandard/ya.make
new file mode 100644
index 00000000000..75cf08f3572
--- /dev/null
+++ b/contrib/python/zstandard/ya.make
@@ -0,0 +1,18 @@
+PY23_LIBRARY()
+
+LICENSE(Service-Py23-Proxy)
+
+IF (PYTHON2)
+    PEERDIR(contrib/python/zstandard/py2)
+ELSE()
+    PEERDIR(contrib/python/zstandard/py3)
+ENDIF()
+
+NO_LINT()
+
+END()
+
+RECURSE(
+    py2
+    py3
+)