// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

#include "parquet/encoding.h"

#include <algorithm>
#include <cstdint>
#include <cstdlib>
#include <limits>
#include <memory>
#include <string>
#include <utility>
#include <vector>

#include "arrow/array.h"
#include "arrow/array/builder_dict.h"
#include "arrow/stl_allocator.h"
#include "arrow/type_traits.h"
#include "arrow/util/bit_run_reader.h"
#include "arrow/util/bit_stream_utils.h"
#include "arrow/util/bit_util.h"
#include "arrow/util/bitmap_ops.h"
#include "arrow/util/bitmap_writer.h"
#include "arrow/util/byte_stream_split.h"
#include "arrow/util/checked_cast.h"
#include "arrow/util/hashing.h"
#include "arrow/util/logging.h"
#include "arrow/util/rle_encoding.h"
#include "arrow/util/ubsan.h"
#include "arrow/visitor_inline.h"

#include "parquet/exception.h"
#include "parquet/platform.h"
#include "parquet/schema.h"
#include "parquet/types.h"

namespace BitUtil = arrow::BitUtil;

using arrow::Status;
using arrow::VisitNullBitmapInline;
using arrow::internal::checked_cast;

template <typename T>
using ArrowPoolVector = std::vector<T, ::arrow::stl::allocator<T>>;

namespace parquet {
namespace {

constexpr int64_t kInMemoryDefaultCapacity = 1024;
// The Parquet spec isn't very clear whether ByteArray lengths are signed or
// unsigned, but the Java implementation uses signed ints.
constexpr size_t kMaxByteArraySize = std::numeric_limits<int32_t>::max();

class EncoderImpl : virtual public Encoder {
 public:
  EncoderImpl(const ColumnDescriptor* descr, Encoding::type encoding, MemoryPool* pool)
      : descr_(descr),
        encoding_(encoding),
        pool_(pool),
        type_length_(descr ? descr->type_length() : -1) {}

  Encoding::type encoding() const override { return encoding_; }

  MemoryPool* memory_pool() const override { return pool_; }

 protected:
  // For accessing type-specific metadata, like FIXED_LEN_BYTE_ARRAY
  const ColumnDescriptor* descr_;
  const Encoding::type encoding_;
  MemoryPool* pool_;

  /// Type length from descr
  int type_length_;
};

// ----------------------------------------------------------------------
// Plain encoder implementation

template <typename DType>
class PlainEncoder : public EncoderImpl, virtual public TypedEncoder<DType> {
 public:
  using T = typename DType::c_type;

  explicit PlainEncoder(const ColumnDescriptor* descr, MemoryPool* pool)
      : EncoderImpl(descr, Encoding::PLAIN, pool), sink_(pool) {}

  int64_t EstimatedDataEncodedSize() override { return sink_.length(); }

  std::shared_ptr<Buffer> FlushValues() override {
    std::shared_ptr<Buffer> buffer;
    PARQUET_THROW_NOT_OK(sink_.Finish(&buffer));
    return buffer;
  }

  using TypedEncoder<DType>::Put;

  void Put(const T* buffer, int num_values) override;

  void Put(const ::arrow::Array& values) override;

  void PutSpaced(const T* src, int num_values, const uint8_t* valid_bits,
                 int64_t valid_bits_offset) override {
    if (valid_bits != NULLPTR) {
      PARQUET_ASSIGN_OR_THROW(auto buffer, ::arrow::AllocateBuffer(num_values * sizeof(T),
                                                                   this->memory_pool()));
      T* data = reinterpret_cast<T*>(buffer->mutable_data());
      int num_valid_values = ::arrow::util::internal::SpacedCompress<T>(
          src, num_values, valid_bits, valid_bits_offset, data);
      Put(data, num_valid_values);
    } else {
      Put(src, num_values);
    }
  }

  void UnsafePutByteArray(const void* data, uint32_t length) {
    DCHECK(length == 0 || data != nullptr) << "Value ptr cannot be NULL";
    sink_.UnsafeAppend(&length, sizeof(uint32_t));
    sink_.UnsafeAppend(data, static_cast<int64_t>(length));
  }

  void Put(const ByteArray& val) {
    // Write the result to the output stream
    const int64_t increment = static_cast<int64_t>(val.len + sizeof(uint32_t));
    if (ARROW_PREDICT_FALSE(sink_.length() + increment > sink_.capacity())) {
      PARQUET_THROW_NOT_OK(sink_.Reserve(increment));
    }
    UnsafePutByteArray(val.ptr, val.len);
  }

 protected:
  template <typename ArrayType>
  void PutBinaryArray(const ArrayType& array) {
    const int64_t total_bytes =
        array.value_offset(array.length()) - array.value_offset(0);
    PARQUET_THROW_NOT_OK(sink_.Reserve(total_bytes + array.length() * sizeof(uint32_t)));

    PARQUET_THROW_NOT_OK(::arrow::VisitArrayDataInline<typename ArrayType::TypeClass>(
        *array.data(),
        [&](::arrow::util::string_view view) {
          if (ARROW_PREDICT_FALSE(view.size() > kMaxByteArraySize)) {
            return Status::Invalid("Parquet cannot store strings with size 2GB or more");
          }
          UnsafePutByteArray(view.data(), static_cast<uint32_t>(view.size()));
          return Status::OK();
        },
        []() { return Status::OK(); }));
  }

  ::arrow::BufferBuilder sink_;
};

template <typename DType>
void PlainEncoder<DType>::Put(const T* buffer, int num_values) {
  if (num_values > 0) {
    PARQUET_THROW_NOT_OK(sink_.Append(buffer, num_values * sizeof(T)));
  }
}

template <>
inline void PlainEncoder<ByteArrayType>::Put(const ByteArray* src, int num_values) {
  for (int i = 0; i < num_values; ++i) {
    Put(src[i]);
  }
}

template <typename ArrayType>
void DirectPutImpl(const ::arrow::Array& values, ::arrow::BufferBuilder* sink) {
  if (values.type_id() != ArrayType::TypeClass::type_id) {
    std::string type_name = ArrayType::TypeClass::type_name();
    throw ParquetException("direct put to " + type_name + " from " +
                           values.type()->ToString() + " not supported");
  }

  using value_type = typename ArrayType::value_type;
  constexpr auto value_size = sizeof(value_type);
  auto raw_values = checked_cast<const ArrayType&>(values).raw_values();

  if (values.null_count() == 0) {
    // no nulls, just dump the data
    PARQUET_THROW_NOT_OK(sink->Append(raw_values, values.length() * value_size));
  } else {
    PARQUET_THROW_NOT_OK(
        sink->Reserve((values.length() - values.null_count()) * value_size));

    for (int64_t i = 0; i < values.length(); i++) {
      if (values.IsValid(i)) {
        sink->UnsafeAppend(&raw_values[i], value_size);
      }
    }
  }
}

template <>
void PlainEncoder<Int32Type>::Put(const ::arrow::Array& values) {
  DirectPutImpl<::arrow::Int32Array>(values, &sink_);
}

template <>
void PlainEncoder<Int64Type>::Put(const ::arrow::Array& values) {
  DirectPutImpl<::arrow::Int64Array>(values, &sink_);
}

template <>
void PlainEncoder<Int96Type>::Put(const ::arrow::Array& values) {
  ParquetException::NYI("direct put to Int96");
}

template <>
void PlainEncoder<FloatType>::Put(const ::arrow::Array& values) {
  DirectPutImpl<::arrow::FloatArray>(values, &sink_);
}

template <>
void PlainEncoder<DoubleType>::Put(const ::arrow::Array& values) {
  DirectPutImpl<::arrow::DoubleArray>(values, &sink_);
}

template <typename DType>
void PlainEncoder<DType>::Put(const ::arrow::Array& values) {
  ParquetException::NYI("direct put of " + values.type()->ToString());
}

void AssertBaseBinary(const ::arrow::Array& values) {
  if (!::arrow::is_base_binary_like(values.type_id())) {
    throw ParquetException("Only BaseBinaryArray and subclasses supported");
  }
}

template <>
inline void PlainEncoder<ByteArrayType>::Put(const ::arrow::Array& values) {
  AssertBaseBinary(values);

  if (::arrow::is_binary_like(values.type_id())) {
    PutBinaryArray(checked_cast<const ::arrow::BinaryArray&>(values));
  } else {
    DCHECK(::arrow::is_large_binary_like(values.type_id()));
    PutBinaryArray(checked_cast<const ::arrow::LargeBinaryArray&>(values));
  }
}

void AssertFixedSizeBinary(const ::arrow::Array& values, int type_length) {
  if (values.type_id() != ::arrow::Type::FIXED_SIZE_BINARY &&
      values.type_id() != ::arrow::Type::DECIMAL) {
    throw ParquetException("Only FixedSizeBinaryArray and subclasses supported");
  }
  if (checked_cast<const ::arrow::FixedSizeBinaryType&>(*values.type()).byte_width() !=
      type_length) {
    throw ParquetException("Size mismatch: " + values.type()->ToString() +
                           " should have been " + std::to_string(type_length) + " wide");
  }
}

template <>
inline void PlainEncoder<FLBAType>::Put(const ::arrow::Array& values) {
  AssertFixedSizeBinary(values, descr_->type_length());
  const auto& data = checked_cast<const ::arrow::FixedSizeBinaryArray&>(values);

  if (data.null_count() == 0) {
    // no nulls, just dump the data
    PARQUET_THROW_NOT_OK(
        sink_.Append(data.raw_values(), data.length() * data.byte_width()));
  } else {
    const int64_t total_bytes =
        data.length() * data.byte_width() - data.null_count() * data.byte_width();
    PARQUET_THROW_NOT_OK(sink_.Reserve(total_bytes));
    for (int64_t i = 0; i < data.length(); i++) {
      if (data.IsValid(i)) {
        sink_.UnsafeAppend(data.Value(i), data.byte_width());
      }
    }
  }
}

template <>
inline void PlainEncoder<FLBAType>::Put(const FixedLenByteArray* src, int num_values) {
  if (descr_->type_length() == 0) {
    return;
  }
  for (int i = 0; i < num_values; ++i) {
    // Write the result to the output stream
    DCHECK(src[i].ptr != nullptr) << "Value ptr cannot be NULL";
    PARQUET_THROW_NOT_OK(sink_.Append(src[i].ptr, descr_->type_length()));
  }
}

template <>
class PlainEncoder<BooleanType> : public EncoderImpl, virtual public BooleanEncoder {
 public:
  explicit PlainEncoder(const ColumnDescriptor* descr, MemoryPool* pool)
      : EncoderImpl(descr, Encoding::PLAIN, pool),
        bits_available_(kInMemoryDefaultCapacity * 8),
        bits_buffer_(AllocateBuffer(pool, kInMemoryDefaultCapacity)),
        sink_(pool),
        bit_writer_(bits_buffer_->mutable_data(),
                    static_cast<int>(bits_buffer_->size())) {}

  int64_t EstimatedDataEncodedSize() override;
  std::shared_ptr<Buffer> FlushValues() override;

  void Put(const bool* src, int num_values) override;

  void Put(const std::vector<bool>& src, int num_values) override;

  void PutSpaced(const bool* src, int num_values, const uint8_t* valid_bits,
                 int64_t valid_bits_offset) override {
    if (valid_bits != NULLPTR) {
      PARQUET_ASSIGN_OR_THROW(auto buffer, ::arrow::AllocateBuffer(num_values * sizeof(T),
                                                                   this->memory_pool()));
      T* data = reinterpret_cast<T*>(buffer->mutable_data());
      int num_valid_values = ::arrow::util::internal::SpacedCompress<T>(
          src, num_values, valid_bits, valid_bits_offset, data);
      Put(data, num_valid_values);
    } else {
      Put(src, num_values);
    }
  }

  void Put(const ::arrow::Array& values) override {
    if (values.type_id() != ::arrow::Type::BOOL) {
      throw ParquetException("direct put to boolean from " + values.type()->ToString() +
                             " not supported");
    }

    const auto& data = checked_cast<const ::arrow::BooleanArray&>(values);
    if (data.null_count() == 0) {
      PARQUET_THROW_NOT_OK(sink_.Reserve(BitUtil::BytesForBits(data.length())));
      // no nulls, just dump the data
      ::arrow::internal::CopyBitmap(data.data()->GetValues<uint8_t>(1), data.offset(),
                                    data.length(), sink_.mutable_data(), sink_.length());
    } else {
      auto n_valid = BitUtil::BytesForBits(data.length() - data.null_count());
      PARQUET_THROW_NOT_OK(sink_.Reserve(n_valid));
      ::arrow::internal::FirstTimeBitmapWriter writer(sink_.mutable_data(),
                                                      sink_.length(), n_valid);

      for (int64_t i = 0; i < data.length(); i++) {
        if (data.IsValid(i)) {
          if (data.Value(i)) {
            writer.Set();
          } else {
            writer.Clear();
          }
          writer.Next();
        }
      }
      writer.Finish();
    }
    sink_.UnsafeAdvance(data.length());
  }

 private:
  int bits_available_;
  std::shared_ptr<ResizableBuffer> bits_buffer_;
  ::arrow::BufferBuilder sink_;
  ::arrow::BitUtil::BitWriter bit_writer_;

  template <typename SequenceType>
  void PutImpl(const SequenceType& src, int num_values);
};

template <typename SequenceType>
void PlainEncoder<BooleanType>::PutImpl(const SequenceType& src, int num_values) {
  int bit_offset = 0;
  if (bits_available_ > 0) {
    int bits_to_write = std::min(bits_available_, num_values);
    for (int i = 0; i < bits_to_write; i++) {
      bit_writer_.PutValue(src[i], 1);
    }
    bits_available_ -= bits_to_write;
    bit_offset = bits_to_write;

    if (bits_available_ == 0) {
      bit_writer_.Flush();
      PARQUET_THROW_NOT_OK(
          sink_.Append(bit_writer_.buffer(), bit_writer_.bytes_written()));
      bit_writer_.Clear();
    }
  }

  int bits_remaining = num_values - bit_offset;
  while (bit_offset < num_values) {
    bits_available_ = static_cast<int>(bits_buffer_->size()) * 8;

    int bits_to_write = std::min(bits_available_, bits_remaining);
    for (int i = bit_offset; i < bit_offset + bits_to_write; i++) {
      bit_writer_.PutValue(src[i], 1);
    }
    bit_offset += bits_to_write;
    bits_available_ -= bits_to_write;
    bits_remaining -= bits_to_write;

    if (bits_available_ == 0) {
      bit_writer_.Flush();
      PARQUET_THROW_NOT_OK(
          sink_.Append(bit_writer_.buffer(), bit_writer_.bytes_written()));
      bit_writer_.Clear();
    }
  }
}

int64_t PlainEncoder<BooleanType>::EstimatedDataEncodedSize() {
  int64_t position = sink_.length();
  return position + bit_writer_.bytes_written();
}

std::shared_ptr<Buffer> PlainEncoder<BooleanType>::FlushValues() {
  if (bits_available_ > 0) {
    bit_writer_.Flush();
    PARQUET_THROW_NOT_OK(sink_.Append(bit_writer_.buffer(), bit_writer_.bytes_written()));
    bit_writer_.Clear();
    bits_available_ = static_cast<int>(bits_buffer_->size()) * 8;
  }

  std::shared_ptr<Buffer> buffer;
  PARQUET_THROW_NOT_OK(sink_.Finish(&buffer));
  return buffer;
}

void PlainEncoder<BooleanType>::Put(const bool* src, int num_values) {
  PutImpl(src, num_values);
}

void PlainEncoder<BooleanType>::Put(const std::vector<bool>& src, int num_values) {
  PutImpl(src, num_values);
}

// ----------------------------------------------------------------------
// DictEncoder<T> implementations

template <typename DType>
struct DictEncoderTraits {
  using c_type = typename DType::c_type;
  using MemoTableType = ::arrow::internal::ScalarMemoTable<c_type>;
};

template <>
struct DictEncoderTraits<ByteArrayType> {
  using MemoTableType = ::arrow::internal::BinaryMemoTable<::arrow::BinaryBuilder>;
};

template <>
struct DictEncoderTraits<FLBAType> {
  using MemoTableType = ::arrow::internal::BinaryMemoTable<::arrow::BinaryBuilder>;
};

// Initially 1024 elements
static constexpr int32_t kInitialHashTableSize = 1 << 10;

/// See the dictionary encoding section of
/// https://github.com/Parquet/parquet-format.  The encoding supports
/// streaming encoding. Values are encoded as they are added while the
/// dictionary is being constructed. At any time, the buffered values
/// can be written out with the current dictionary size. More values
/// can then be added to the encoder, including new dictionary
/// entries.
template <typename DType>
class DictEncoderImpl : public EncoderImpl, virtual public DictEncoder<DType> {
  using MemoTableType = typename DictEncoderTraits<DType>::MemoTableType;

 public:
  typedef typename DType::c_type T;

  explicit DictEncoderImpl(const ColumnDescriptor* desc, MemoryPool* pool)
      : EncoderImpl(desc, Encoding::PLAIN_DICTIONARY, pool),
        buffered_indices_(::arrow::stl::allocator<int32_t>(pool)),
        dict_encoded_size_(0),
        memo_table_(pool, kInitialHashTableSize) {}

  ~DictEncoderImpl() override { DCHECK(buffered_indices_.empty()); }

  int dict_encoded_size() override { return dict_encoded_size_; }

  int WriteIndices(uint8_t* buffer, int buffer_len) override {
    // Write bit width in first byte
    *buffer = static_cast<uint8_t>(bit_width());
    ++buffer;
    --buffer_len;

    ::arrow::util::RleEncoder encoder(buffer, buffer_len, bit_width());

    for (int32_t index : buffered_indices_) {
      if (!encoder.Put(index)) return -1;
    }
    encoder.Flush();

    ClearIndices();
    return 1 + encoder.len();
  }

  void set_type_length(int type_length) { this->type_length_ = type_length; }

  /// Returns a conservative estimate of the number of bytes needed to encode the buffered
  /// indices. Used to size the buffer passed to WriteIndices().
  int64_t EstimatedDataEncodedSize() override {
    // Note: because of the way RleEncoder::CheckBufferFull() is called, we have to
    // reserve
    // an extra "RleEncoder::MinBufferSize" bytes. These extra bytes won't be used
    // but not reserving them would cause the encoder to fail.
    return 1 +
           ::arrow::util::RleEncoder::MaxBufferSize(
               bit_width(), static_cast<int>(buffered_indices_.size())) +
           ::arrow::util::RleEncoder::MinBufferSize(bit_width());
  }

  /// The minimum bit width required to encode the currently buffered indices.
  int bit_width() const override {
    if (ARROW_PREDICT_FALSE(num_entries() == 0)) return 0;
    if (ARROW_PREDICT_FALSE(num_entries() == 1)) return 1;
    return BitUtil::Log2(num_entries());
  }

  /// Encode value. Note that this does not actually write any data, just
  /// buffers the value's index to be written later.
  inline void Put(const T& value);

  // Not implemented for other data types
  inline void PutByteArray(const void* ptr, int32_t length);

  void Put(const T* src, int num_values) override {
    for (int32_t i = 0; i < num_values; i++) {
      Put(src[i]);
    }
  }

  void PutSpaced(const T* src, int num_values, const uint8_t* valid_bits,
                 int64_t valid_bits_offset) override {
    ::arrow::internal::VisitSetBitRunsVoid(valid_bits, valid_bits_offset, num_values,
                                           [&](int64_t position, int64_t length) {
                                             for (int64_t i = 0; i < length; i++) {
                                               Put(src[i + position]);
                                             }
                                           });
  }

  using TypedEncoder<DType>::Put;

  void Put(const ::arrow::Array& values) override;
  void PutDictionary(const ::arrow::Array& values) override;

  template <typename ArrowType, typename T = typename ArrowType::c_type>
  void PutIndicesTyped(const ::arrow::Array& data) {
    auto values = data.data()->GetValues<T>(1);
    size_t buffer_position = buffered_indices_.size();
    buffered_indices_.resize(buffer_position +
                             static_cast<size_t>(data.length() - data.null_count()));
    ::arrow::internal::VisitSetBitRunsVoid(
        data.null_bitmap_data(), data.offset(), data.length(),
        [&](int64_t position, int64_t length) {
          for (int64_t i = 0; i < length; ++i) {
            buffered_indices_[buffer_position++] =
                static_cast<int32_t>(values[i + position]);
          }
        });
  }

  void PutIndices(const ::arrow::Array& data) override {
    switch (data.type()->id()) {
      case ::arrow::Type::UINT8:
      case ::arrow::Type::INT8:
        return PutIndicesTyped<::arrow::UInt8Type>(data);
      case ::arrow::Type::UINT16:
      case ::arrow::Type::INT16:
        return PutIndicesTyped<::arrow::UInt16Type>(data);
      case ::arrow::Type::UINT32:
      case ::arrow::Type::INT32:
        return PutIndicesTyped<::arrow::UInt32Type>(data);
      case ::arrow::Type::UINT64:
      case ::arrow::Type::INT64:
        return PutIndicesTyped<::arrow::UInt64Type>(data);
      default:
        throw ParquetException("Passed non-integer array to PutIndices");
    }
  }

  std::shared_ptr<Buffer> FlushValues() override {
    std::shared_ptr<ResizableBuffer> buffer =
        AllocateBuffer(this->pool_, EstimatedDataEncodedSize());
    int result_size = WriteIndices(buffer->mutable_data(),
                                   static_cast<int>(EstimatedDataEncodedSize()));
    PARQUET_THROW_NOT_OK(buffer->Resize(result_size, false));
    return std::move(buffer);
  }

  /// Writes out the encoded dictionary to buffer. buffer must be preallocated to
  /// dict_encoded_size() bytes.
  void WriteDict(uint8_t* buffer) override;

  /// The number of entries in the dictionary.
  int num_entries() const override { return memo_table_.size(); }

 private:
  /// Clears all the indices (but leaves the dictionary).
  void ClearIndices() { buffered_indices_.clear(); }

  /// Indices that have not yet be written out by WriteIndices().
  ArrowPoolVector<int32_t> buffered_indices_;

  template <typename ArrayType>
  void PutBinaryArray(const ArrayType& array) {
    PARQUET_THROW_NOT_OK(::arrow::VisitArrayDataInline<typename ArrayType::TypeClass>(
        *array.data(),
        [&](::arrow::util::string_view view) {
          if (ARROW_PREDICT_FALSE(view.size() > kMaxByteArraySize)) {
            return Status::Invalid("Parquet cannot store strings with size 2GB or more");
          }
          PutByteArray(view.data(), static_cast<uint32_t>(view.size()));
          return Status::OK();
        },
        []() { return Status::OK(); }));
  }

  template <typename ArrayType>
  void PutBinaryDictionaryArray(const ArrayType& array) {
    DCHECK_EQ(array.null_count(), 0);
    for (int64_t i = 0; i < array.length(); i++) {
      auto v = array.GetView(i);
      if (ARROW_PREDICT_FALSE(v.size() > kMaxByteArraySize)) {
        throw ParquetException("Parquet cannot store strings with size 2GB or more");
      }
      dict_encoded_size_ += static_cast<int>(v.size() + sizeof(uint32_t));
      int32_t unused_memo_index;
      PARQUET_THROW_NOT_OK(memo_table_.GetOrInsert(
          v.data(), static_cast<int32_t>(v.size()), &unused_memo_index));
    }
  }

  /// The number of bytes needed to encode the dictionary.
  int dict_encoded_size_;

  MemoTableType memo_table_;
};

template <typename DType>
void DictEncoderImpl<DType>::WriteDict(uint8_t* buffer) {
  // For primitive types, only a memcpy
  DCHECK_EQ(static_cast<size_t>(dict_encoded_size_), sizeof(T) * memo_table_.size());
  memo_table_.CopyValues(0 /* start_pos */, reinterpret_cast<T*>(buffer));
}

// ByteArray and FLBA already have the dictionary encoded in their data heaps
template <>
void DictEncoderImpl<ByteArrayType>::WriteDict(uint8_t* buffer) {
  memo_table_.VisitValues(0, [&buffer](const ::arrow::util::string_view& v) {
    uint32_t len = static_cast<uint32_t>(v.length());
    memcpy(buffer, &len, sizeof(len));
    buffer += sizeof(len);
    memcpy(buffer, v.data(), len);
    buffer += len;
  });
}

template <>
void DictEncoderImpl<FLBAType>::WriteDict(uint8_t* buffer) {
  memo_table_.VisitValues(0, [&](const ::arrow::util::string_view& v) {
    DCHECK_EQ(v.length(), static_cast<size_t>(type_length_));
    memcpy(buffer, v.data(), type_length_);
    buffer += type_length_;
  });
}

template <typename DType>
inline void DictEncoderImpl<DType>::Put(const T& v) {
  // Put() implementation for primitive types
  auto on_found = [](int32_t memo_index) {};
  auto on_not_found = [this](int32_t memo_index) {
    dict_encoded_size_ += static_cast<int>(sizeof(T));
  };

  int32_t memo_index;
  PARQUET_THROW_NOT_OK(memo_table_.GetOrInsert(v, on_found, on_not_found, &memo_index));
  buffered_indices_.push_back(memo_index);
}

template <typename DType>
inline void DictEncoderImpl<DType>::PutByteArray(const void* ptr, int32_t length) {
  DCHECK(false);
}

template <>
inline void DictEncoderImpl<ByteArrayType>::PutByteArray(const void* ptr,
                                                         int32_t length) {
  static const uint8_t empty[] = {0};

  auto on_found = [](int32_t memo_index) {};
  auto on_not_found = [&](int32_t memo_index) {
    dict_encoded_size_ += static_cast<int>(length + sizeof(uint32_t));
  };

  DCHECK(ptr != nullptr || length == 0);
  ptr = (ptr != nullptr) ? ptr : empty;
  int32_t memo_index;
  PARQUET_THROW_NOT_OK(
      memo_table_.GetOrInsert(ptr, length, on_found, on_not_found, &memo_index));
  buffered_indices_.push_back(memo_index);
}

template <>
inline void DictEncoderImpl<ByteArrayType>::Put(const ByteArray& val) {
  return PutByteArray(val.ptr, static_cast<int32_t>(val.len));
}

template <>
inline void DictEncoderImpl<FLBAType>::Put(const FixedLenByteArray& v) {
  static const uint8_t empty[] = {0};

  auto on_found = [](int32_t memo_index) {};
  auto on_not_found = [this](int32_t memo_index) { dict_encoded_size_ += type_length_; };

  DCHECK(v.ptr != nullptr || type_length_ == 0);
  const void* ptr = (v.ptr != nullptr) ? v.ptr : empty;
  int32_t memo_index;
  PARQUET_THROW_NOT_OK(
      memo_table_.GetOrInsert(ptr, type_length_, on_found, on_not_found, &memo_index));
  buffered_indices_.push_back(memo_index);
}

template <>
void DictEncoderImpl<Int96Type>::Put(const ::arrow::Array& values) {
  ParquetException::NYI("Direct put to Int96");
}

template <>
void DictEncoderImpl<Int96Type>::PutDictionary(const ::arrow::Array& values) {
  ParquetException::NYI("Direct put to Int96");
}

template <typename DType>
void DictEncoderImpl<DType>::Put(const ::arrow::Array& values) {
  using ArrayType = typename ::arrow::CTypeTraits<typename DType::c_type>::ArrayType;
  const auto& data = checked_cast<const ArrayType&>(values);
  if (data.null_count() == 0) {
    // no nulls, just dump the data
    for (int64_t i = 0; i < data.length(); i++) {
      Put(data.Value(i));
    }
  } else {
    for (int64_t i = 0; i < data.length(); i++) {
      if (data.IsValid(i)) {
        Put(data.Value(i));
      }
    }
  }
}

template <>
void DictEncoderImpl<FLBAType>::Put(const ::arrow::Array& values) {
  AssertFixedSizeBinary(values, type_length_);
  const auto& data = checked_cast<const ::arrow::FixedSizeBinaryArray&>(values);
  if (data.null_count() == 0) {
    // no nulls, just dump the data
    for (int64_t i = 0; i < data.length(); i++) {
      Put(FixedLenByteArray(data.Value(i)));
    }
  } else {
    std::vector<uint8_t> empty(type_length_, 0);
    for (int64_t i = 0; i < data.length(); i++) {
      if (data.IsValid(i)) {
        Put(FixedLenByteArray(data.Value(i)));
      }
    }
  }
}

template <>
void DictEncoderImpl<ByteArrayType>::Put(const ::arrow::Array& values) {
  AssertBaseBinary(values);
  if (::arrow::is_binary_like(values.type_id())) {
    PutBinaryArray(checked_cast<const ::arrow::BinaryArray&>(values));
  } else {
    DCHECK(::arrow::is_large_binary_like(values.type_id()));
    PutBinaryArray(checked_cast<const ::arrow::LargeBinaryArray&>(values));
  }
}

template <typename DType>
void AssertCanPutDictionary(DictEncoderImpl<DType>* encoder, const ::arrow::Array& dict) {
  if (dict.null_count() > 0) {
    throw ParquetException("Inserted dictionary cannot cannot contain nulls");
  }

  if (encoder->num_entries() > 0) {
    throw ParquetException("Can only call PutDictionary on an empty DictEncoder");
  }
}

template <typename DType>
void DictEncoderImpl<DType>::PutDictionary(const ::arrow::Array& values) {
  AssertCanPutDictionary(this, values);

  using ArrayType = typename ::arrow::CTypeTraits<typename DType::c_type>::ArrayType;
  const auto& data = checked_cast<const ArrayType&>(values);

  dict_encoded_size_ += static_cast<int>(sizeof(typename DType::c_type) * data.length());
  for (int64_t i = 0; i < data.length(); i++) {
    int32_t unused_memo_index;
    PARQUET_THROW_NOT_OK(memo_table_.GetOrInsert(data.Value(i), &unused_memo_index));
  }
}

template <>
void DictEncoderImpl<FLBAType>::PutDictionary(const ::arrow::Array& values) {
  AssertFixedSizeBinary(values, type_length_);
  AssertCanPutDictionary(this, values);

  const auto& data = checked_cast<const ::arrow::FixedSizeBinaryArray&>(values);

  dict_encoded_size_ += static_cast<int>(type_length_ * data.length());
  for (int64_t i = 0; i < data.length(); i++) {
    int32_t unused_memo_index;
    PARQUET_THROW_NOT_OK(
        memo_table_.GetOrInsert(data.Value(i), type_length_, &unused_memo_index));
  }
}

template <>
void DictEncoderImpl<ByteArrayType>::PutDictionary(const ::arrow::Array& values) {
  AssertBaseBinary(values);
  AssertCanPutDictionary(this, values);

  if (::arrow::is_binary_like(values.type_id())) {
    PutBinaryDictionaryArray(checked_cast<const ::arrow::BinaryArray&>(values));
  } else {
    DCHECK(::arrow::is_large_binary_like(values.type_id()));
    PutBinaryDictionaryArray(checked_cast<const ::arrow::LargeBinaryArray&>(values));
  }
}

// ----------------------------------------------------------------------
// ByteStreamSplitEncoder<T> implementations

template <typename DType>
class ByteStreamSplitEncoder : public EncoderImpl, virtual public TypedEncoder<DType> {
 public:
  using T = typename DType::c_type;
  using TypedEncoder<DType>::Put;

  explicit ByteStreamSplitEncoder(
      const ColumnDescriptor* descr,
      ::arrow::MemoryPool* pool = ::arrow::default_memory_pool());

  int64_t EstimatedDataEncodedSize() override;
  std::shared_ptr<Buffer> FlushValues() override;

  void Put(const T* buffer, int num_values) override;
  void Put(const ::arrow::Array& values) override;
  void PutSpaced(const T* src, int num_values, const uint8_t* valid_bits,
                 int64_t valid_bits_offset) override;

 protected:
  template <typename ArrowType>
  void PutImpl(const ::arrow::Array& values) {
    if (values.type_id() != ArrowType::type_id) {
      throw ParquetException(std::string() + "direct put to " + ArrowType::type_name() +
                             " from " + values.type()->ToString() + " not supported");
    }
    const auto& data = *values.data();
    PutSpaced(data.GetValues<typename ArrowType::c_type>(1),
              static_cast<int>(data.length), data.GetValues<uint8_t>(0, 0), data.offset);
  }

  ::arrow::BufferBuilder sink_;
  int64_t num_values_in_buffer_;
};

template <typename DType>
ByteStreamSplitEncoder<DType>::ByteStreamSplitEncoder(const ColumnDescriptor* descr,
                                                      ::arrow::MemoryPool* pool)
    : EncoderImpl(descr, Encoding::BYTE_STREAM_SPLIT, pool),
      sink_{pool},
      num_values_in_buffer_{0} {}

template <typename DType>
int64_t ByteStreamSplitEncoder<DType>::EstimatedDataEncodedSize() {
  return sink_.length();
}

template <typename DType>
std::shared_ptr<Buffer> ByteStreamSplitEncoder<DType>::FlushValues() {
  std::shared_ptr<ResizableBuffer> output_buffer =
      AllocateBuffer(this->memory_pool(), EstimatedDataEncodedSize());
  uint8_t* output_buffer_raw = output_buffer->mutable_data();
  const uint8_t* raw_values = sink_.data();
  ::arrow::util::internal::ByteStreamSplitEncode<T>(raw_values, num_values_in_buffer_,
                                                    output_buffer_raw);
  sink_.Reset();
  num_values_in_buffer_ = 0;
  return std::move(output_buffer);
}

template <typename DType>
void ByteStreamSplitEncoder<DType>::Put(const T* buffer, int num_values) {
  if (num_values > 0) {
    PARQUET_THROW_NOT_OK(sink_.Append(buffer, num_values * sizeof(T)));
    num_values_in_buffer_ += num_values;
  }
}

template <>
void ByteStreamSplitEncoder<FloatType>::Put(const ::arrow::Array& values) {
  PutImpl<::arrow::FloatType>(values);
}

template <>
void ByteStreamSplitEncoder<DoubleType>::Put(const ::arrow::Array& values) {
  PutImpl<::arrow::DoubleType>(values);
}

template <typename DType>
void ByteStreamSplitEncoder<DType>::PutSpaced(const T* src, int num_values,
                                              const uint8_t* valid_bits,
                                              int64_t valid_bits_offset) {
  if (valid_bits != NULLPTR) {
    PARQUET_ASSIGN_OR_THROW(auto buffer, ::arrow::AllocateBuffer(num_values * sizeof(T),
                                                                 this->memory_pool()));
    T* data = reinterpret_cast<T*>(buffer->mutable_data());
    int num_valid_values = ::arrow::util::internal::SpacedCompress<T>(
        src, num_values, valid_bits, valid_bits_offset, data);
    Put(data, num_valid_values);
  } else {
    Put(src, num_values);
  }
}

class DecoderImpl : virtual public Decoder {
 public:
  void SetData(int num_values, const uint8_t* data, int len) override {
    num_values_ = num_values;
    data_ = data;
    len_ = len;
  }

  int values_left() const override { return num_values_; }
  Encoding::type encoding() const override { return encoding_; }

 protected:
  explicit DecoderImpl(const ColumnDescriptor* descr, Encoding::type encoding)
      : descr_(descr), encoding_(encoding), num_values_(0), data_(NULLPTR), len_(0) {}

  // For accessing type-specific metadata, like FIXED_LEN_BYTE_ARRAY
  const ColumnDescriptor* descr_;

  const Encoding::type encoding_;
  int num_values_;
  const uint8_t* data_;
  int len_;
  int type_length_;
};

template <typename DType>
class PlainDecoder : public DecoderImpl, virtual public TypedDecoder<DType> {
 public:
  using T = typename DType::c_type;
  explicit PlainDecoder(const ColumnDescriptor* descr);

  int Decode(T* buffer, int max_values) override;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<DType>::Accumulator* builder) override;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<DType>::DictAccumulator* builder) override;
};

template <>
inline int PlainDecoder<Int96Type>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<Int96Type>::Accumulator* builder) {
  ParquetException::NYI("DecodeArrow not supported for Int96");
}

template <>
inline int PlainDecoder<Int96Type>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<Int96Type>::DictAccumulator* builder) {
  ParquetException::NYI("DecodeArrow not supported for Int96");
}

template <>
inline int PlainDecoder<BooleanType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<BooleanType>::DictAccumulator* builder) {
  ParquetException::NYI("dictionaries of BooleanType");
}

template <typename DType>
int PlainDecoder<DType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<DType>::Accumulator* builder) {
  using value_type = typename DType::c_type;

  constexpr int value_size = static_cast<int>(sizeof(value_type));
  int values_decoded = num_values - null_count;
  if (ARROW_PREDICT_FALSE(len_ < value_size * values_decoded)) {
    ParquetException::EofException();
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        builder->UnsafeAppend(::arrow::util::SafeLoadAs<value_type>(data_));
        data_ += sizeof(value_type);
      },
      [&]() { builder->UnsafeAppendNull(); });

  num_values_ -= values_decoded;
  len_ -= sizeof(value_type) * values_decoded;
  return values_decoded;
}

template <typename DType>
int PlainDecoder<DType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<DType>::DictAccumulator* builder) {
  using value_type = typename DType::c_type;

  constexpr int value_size = static_cast<int>(sizeof(value_type));
  int values_decoded = num_values - null_count;
  if (ARROW_PREDICT_FALSE(len_ < value_size * values_decoded)) {
    ParquetException::EofException();
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        PARQUET_THROW_NOT_OK(
            builder->Append(::arrow::util::SafeLoadAs<value_type>(data_)));
        data_ += sizeof(value_type);
      },
      [&]() { PARQUET_THROW_NOT_OK(builder->AppendNull()); });

  num_values_ -= values_decoded;
  len_ -= sizeof(value_type) * values_decoded;
  return values_decoded;
}

// Decode routine templated on C++ type rather than type enum
template <typename T>
inline int DecodePlain(const uint8_t* data, int64_t data_size, int num_values,
                       int type_length, T* out) {
  int64_t bytes_to_decode = num_values * static_cast<int64_t>(sizeof(T));
  if (bytes_to_decode > data_size || bytes_to_decode > INT_MAX) {
    ParquetException::EofException();
  }
  // If bytes_to_decode == 0, data could be null
  if (bytes_to_decode > 0) {
    memcpy(out, data, bytes_to_decode);
  }
  return static_cast<int>(bytes_to_decode);
}

template <typename DType>
PlainDecoder<DType>::PlainDecoder(const ColumnDescriptor* descr)
    : DecoderImpl(descr, Encoding::PLAIN) {
  if (descr_ && descr_->physical_type() == Type::FIXED_LEN_BYTE_ARRAY) {
    type_length_ = descr_->type_length();
  } else {
    type_length_ = -1;
  }
}

// Template specialization for BYTE_ARRAY. The written values do not own their
// own data.

static inline int64_t ReadByteArray(const uint8_t* data, int64_t data_size,
                                    ByteArray* out) {
  if (ARROW_PREDICT_FALSE(data_size < 4)) {
    ParquetException::EofException();
  }
  const int32_t len = ::arrow::util::SafeLoadAs<int32_t>(data);
  if (len < 0) {
    throw ParquetException("Invalid BYTE_ARRAY value");
  }
  const int64_t consumed_length = static_cast<int64_t>(len) + 4;
  if (ARROW_PREDICT_FALSE(data_size < consumed_length)) {
    ParquetException::EofException();
  }
  *out = ByteArray{static_cast<uint32_t>(len), data + 4};
  return consumed_length;
}

template <>
inline int DecodePlain<ByteArray>(const uint8_t* data, int64_t data_size, int num_values,
                                  int type_length, ByteArray* out) {
  int bytes_decoded = 0;
  for (int i = 0; i < num_values; ++i) {
    const auto increment = ReadByteArray(data, data_size, out + i);
    if (ARROW_PREDICT_FALSE(increment > INT_MAX - bytes_decoded)) {
      throw ParquetException("BYTE_ARRAY chunk too large");
    }
    data += increment;
    data_size -= increment;
    bytes_decoded += static_cast<int>(increment);
  }
  return bytes_decoded;
}

// Template specialization for FIXED_LEN_BYTE_ARRAY. The written values do not
// own their own data.
template <>
inline int DecodePlain<FixedLenByteArray>(const uint8_t* data, int64_t data_size,
                                          int num_values, int type_length,
                                          FixedLenByteArray* out) {
  int64_t bytes_to_decode = static_cast<int64_t>(type_length) * num_values;
  if (bytes_to_decode > data_size || bytes_to_decode > INT_MAX) {
    ParquetException::EofException();
  }
  for (int i = 0; i < num_values; ++i) {
    out[i].ptr = data;
    data += type_length;
    data_size -= type_length;
  }
  return static_cast<int>(bytes_to_decode);
}

template <typename DType>
int PlainDecoder<DType>::Decode(T* buffer, int max_values) {
  max_values = std::min(max_values, num_values_);
  int bytes_consumed = DecodePlain<T>(data_, len_, max_values, type_length_, buffer);
  data_ += bytes_consumed;
  len_ -= bytes_consumed;
  num_values_ -= max_values;
  return max_values;
}

class PlainBooleanDecoder : public DecoderImpl,
                            virtual public TypedDecoder<BooleanType>,
                            virtual public BooleanDecoder {
 public:
  explicit PlainBooleanDecoder(const ColumnDescriptor* descr);
  void SetData(int num_values, const uint8_t* data, int len) override;

  // Two flavors of bool decoding
  int Decode(uint8_t* buffer, int max_values) override;
  int Decode(bool* buffer, int max_values) override;
  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<BooleanType>::Accumulator* out) override;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<BooleanType>::DictAccumulator* out) override;

 private:
  std::unique_ptr<::arrow::BitUtil::BitReader> bit_reader_;
};

PlainBooleanDecoder::PlainBooleanDecoder(const ColumnDescriptor* descr)
    : DecoderImpl(descr, Encoding::PLAIN) {}

void PlainBooleanDecoder::SetData(int num_values, const uint8_t* data, int len) {
  num_values_ = num_values;
  bit_reader_.reset(new BitUtil::BitReader(data, len));
}

int PlainBooleanDecoder::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<BooleanType>::Accumulator* builder) {
  int values_decoded = num_values - null_count;
  if (ARROW_PREDICT_FALSE(num_values_ < values_decoded)) {
    ParquetException::EofException();
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        bool value;
        ARROW_IGNORE_EXPR(bit_reader_->GetValue(1, &value));
        builder->UnsafeAppend(value);
      },
      [&]() { builder->UnsafeAppendNull(); });

  num_values_ -= values_decoded;
  return values_decoded;
}

inline int PlainBooleanDecoder::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<BooleanType>::DictAccumulator* builder) {
  ParquetException::NYI("dictionaries of BooleanType");
}

int PlainBooleanDecoder::Decode(uint8_t* buffer, int max_values) {
  max_values = std::min(max_values, num_values_);
  bool val;
  ::arrow::internal::BitmapWriter bit_writer(buffer, 0, max_values);
  for (int i = 0; i < max_values; ++i) {
    if (!bit_reader_->GetValue(1, &val)) {
      ParquetException::EofException();
    }
    if (val) {
      bit_writer.Set();
    }
    bit_writer.Next();
  }
  bit_writer.Finish();
  num_values_ -= max_values;
  return max_values;
}

int PlainBooleanDecoder::Decode(bool* buffer, int max_values) {
  max_values = std::min(max_values, num_values_);
  if (bit_reader_->GetBatch(1, buffer, max_values) != max_values) {
    ParquetException::EofException();
  }
  num_values_ -= max_values;
  return max_values;
}

struct ArrowBinaryHelper {
  explicit ArrowBinaryHelper(typename EncodingTraits<ByteArrayType>::Accumulator* out) {
    this->out = out;
    this->builder = out->builder.get();
    this->chunk_space_remaining =
        ::arrow::kBinaryMemoryLimit - this->builder->value_data_length();
  }

  Status PushChunk() {
    std::shared_ptr<::arrow::Array> result;
    RETURN_NOT_OK(builder->Finish(&result));
    out->chunks.push_back(result);
    chunk_space_remaining = ::arrow::kBinaryMemoryLimit;
    return Status::OK();
  }

  bool CanFit(int64_t length) const { return length <= chunk_space_remaining; }

  void UnsafeAppend(const uint8_t* data, int32_t length) {
    chunk_space_remaining -= length;
    builder->UnsafeAppend(data, length);
  }

  void UnsafeAppendNull() { builder->UnsafeAppendNull(); }

  Status Append(const uint8_t* data, int32_t length) {
    chunk_space_remaining -= length;
    return builder->Append(data, length);
  }

  Status AppendNull() { return builder->AppendNull(); }

  typename EncodingTraits<ByteArrayType>::Accumulator* out;
  ::arrow::BinaryBuilder* builder;
  int64_t chunk_space_remaining;
};

template <>
inline int PlainDecoder<ByteArrayType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<ByteArrayType>::Accumulator* builder) {
  ParquetException::NYI();
}

template <>
inline int PlainDecoder<ByteArrayType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<ByteArrayType>::DictAccumulator* builder) {
  ParquetException::NYI();
}

template <>
inline int PlainDecoder<FLBAType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<FLBAType>::Accumulator* builder) {
  int values_decoded = num_values - null_count;
  if (ARROW_PREDICT_FALSE(len_ < descr_->type_length() * values_decoded)) {
    ParquetException::EofException();
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        builder->UnsafeAppend(data_);
        data_ += descr_->type_length();
      },
      [&]() { builder->UnsafeAppendNull(); });

  num_values_ -= values_decoded;
  len_ -= descr_->type_length() * values_decoded;
  return values_decoded;
}

template <>
inline int PlainDecoder<FLBAType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<FLBAType>::DictAccumulator* builder) {
  int values_decoded = num_values - null_count;
  if (ARROW_PREDICT_FALSE(len_ < descr_->type_length() * values_decoded)) {
    ParquetException::EofException();
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        PARQUET_THROW_NOT_OK(builder->Append(data_));
        data_ += descr_->type_length();
      },
      [&]() { PARQUET_THROW_NOT_OK(builder->AppendNull()); });

  num_values_ -= values_decoded;
  len_ -= descr_->type_length() * values_decoded;
  return values_decoded;
}

class PlainByteArrayDecoder : public PlainDecoder<ByteArrayType>,
                              virtual public ByteArrayDecoder {
 public:
  using Base = PlainDecoder<ByteArrayType>;
  using Base::DecodeSpaced;
  using Base::PlainDecoder;

  // ----------------------------------------------------------------------
  // Dictionary read paths

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  ::arrow::BinaryDictionary32Builder* builder) override {
    int result = 0;
    PARQUET_THROW_NOT_OK(DecodeArrow(num_values, null_count, valid_bits,
                                     valid_bits_offset, builder, &result));
    return result;
  }

  // ----------------------------------------------------------------------
  // Optimized dense binary read paths

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<ByteArrayType>::Accumulator* out) override {
    int result = 0;
    PARQUET_THROW_NOT_OK(DecodeArrowDense(num_values, null_count, valid_bits,
                                          valid_bits_offset, out, &result));
    return result;
  }

 private:
  Status DecodeArrowDense(int num_values, int null_count, const uint8_t* valid_bits,
                          int64_t valid_bits_offset,
                          typename EncodingTraits<ByteArrayType>::Accumulator* out,
                          int* out_values_decoded) {
    ArrowBinaryHelper helper(out);
    int values_decoded = 0;

    RETURN_NOT_OK(helper.builder->Reserve(num_values));
    RETURN_NOT_OK(helper.builder->ReserveData(
        std::min<int64_t>(len_, helper.chunk_space_remaining)));

    int i = 0;
    RETURN_NOT_OK(VisitNullBitmapInline(
        valid_bits, valid_bits_offset, num_values, null_count,
        [&]() {
          if (ARROW_PREDICT_FALSE(len_ < 4)) {
            ParquetException::EofException();
          }
          auto value_len = ::arrow::util::SafeLoadAs<int32_t>(data_);
          if (ARROW_PREDICT_FALSE(value_len < 0 || value_len > INT32_MAX - 4)) {
            return Status::Invalid("Invalid or corrupted value_len '", value_len, "'");
          }
          auto increment = value_len + 4;
          if (ARROW_PREDICT_FALSE(len_ < increment)) {
            ParquetException::EofException();
          }
          if (ARROW_PREDICT_FALSE(!helper.CanFit(value_len))) {
            // This element would exceed the capacity of a chunk
            RETURN_NOT_OK(helper.PushChunk());
            RETURN_NOT_OK(helper.builder->Reserve(num_values - i));
            RETURN_NOT_OK(helper.builder->ReserveData(
                std::min<int64_t>(len_, helper.chunk_space_remaining)));
          }
          helper.UnsafeAppend(data_ + 4, value_len);
          data_ += increment;
          len_ -= increment;
          ++values_decoded;
          ++i;
          return Status::OK();
        },
        [&]() {
          helper.UnsafeAppendNull();
          ++i;
          return Status::OK();
        }));

    num_values_ -= values_decoded;
    *out_values_decoded = values_decoded;
    return Status::OK();
  }

  template <typename BuilderType>
  Status DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                     int64_t valid_bits_offset, BuilderType* builder,
                     int* out_values_decoded) {
    RETURN_NOT_OK(builder->Reserve(num_values));
    int values_decoded = 0;

    RETURN_NOT_OK(VisitNullBitmapInline(
        valid_bits, valid_bits_offset, num_values, null_count,
        [&]() {
          if (ARROW_PREDICT_FALSE(len_ < 4)) {
            ParquetException::EofException();
          }
          auto value_len = ::arrow::util::SafeLoadAs<int32_t>(data_);
          if (ARROW_PREDICT_FALSE(value_len < 0 || value_len > INT32_MAX - 4)) {
            return Status::Invalid("Invalid or corrupted value_len '", value_len, "'");
          }
          auto increment = value_len + 4;
          if (ARROW_PREDICT_FALSE(len_ < increment)) {
            ParquetException::EofException();
          }
          RETURN_NOT_OK(builder->Append(data_ + 4, value_len));
          data_ += increment;
          len_ -= increment;
          ++values_decoded;
          return Status::OK();
        },
        [&]() { return builder->AppendNull(); }));

    num_values_ -= values_decoded;
    *out_values_decoded = values_decoded;
    return Status::OK();
  }
};

class PlainFLBADecoder : public PlainDecoder<FLBAType>, virtual public FLBADecoder {
 public:
  using Base = PlainDecoder<FLBAType>;
  using Base::PlainDecoder;
};

// ----------------------------------------------------------------------
// Dictionary encoding and decoding

template <typename Type>
class DictDecoderImpl : public DecoderImpl, virtual public DictDecoder<Type> {
 public:
  typedef typename Type::c_type T;

  // Initializes the dictionary with values from 'dictionary'. The data in
  // dictionary is not guaranteed to persist in memory after this call so the
  // dictionary decoder needs to copy the data out if necessary.
  explicit DictDecoderImpl(const ColumnDescriptor* descr,
                           MemoryPool* pool = ::arrow::default_memory_pool())
      : DecoderImpl(descr, Encoding::RLE_DICTIONARY),
        dictionary_(AllocateBuffer(pool, 0)),
        dictionary_length_(0),
        byte_array_data_(AllocateBuffer(pool, 0)),
        byte_array_offsets_(AllocateBuffer(pool, 0)),
        indices_scratch_space_(AllocateBuffer(pool, 0)) {}

  // Perform type-specific initiatialization
  void SetDict(TypedDecoder<Type>* dictionary) override;

  void SetData(int num_values, const uint8_t* data, int len) override {
    num_values_ = num_values;
    if (len == 0) {
      // Initialize dummy decoder to avoid crashes later on
      idx_decoder_ = ::arrow::util::RleDecoder(data, len, /*bit_width=*/1);
      return;
    }
    uint8_t bit_width = *data;
    if (ARROW_PREDICT_FALSE(bit_width >= 64)) {
      throw ParquetException("Invalid or corrupted bit_width");
    }
    idx_decoder_ = ::arrow::util::RleDecoder(++data, --len, bit_width);
  }

  int Decode(T* buffer, int num_values) override {
    num_values = std::min(num_values, num_values_);
    int decoded_values =
        idx_decoder_.GetBatchWithDict(reinterpret_cast<const T*>(dictionary_->data()),
                                      dictionary_length_, buffer, num_values);
    if (decoded_values != num_values) {
      ParquetException::EofException();
    }
    num_values_ -= num_values;
    return num_values;
  }

  int DecodeSpaced(T* buffer, int num_values, int null_count, const uint8_t* valid_bits,
                   int64_t valid_bits_offset) override {
    num_values = std::min(num_values, num_values_);
    if (num_values != idx_decoder_.GetBatchWithDictSpaced(
                          reinterpret_cast<const T*>(dictionary_->data()),
                          dictionary_length_, buffer, num_values, null_count, valid_bits,
                          valid_bits_offset)) {
      ParquetException::EofException();
    }
    num_values_ -= num_values;
    return num_values;
  }

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<Type>::Accumulator* out) override;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<Type>::DictAccumulator* out) override;

  void InsertDictionary(::arrow::ArrayBuilder* builder) override;

  int DecodeIndicesSpaced(int num_values, int null_count, const uint8_t* valid_bits,
                          int64_t valid_bits_offset,
                          ::arrow::ArrayBuilder* builder) override {
    if (num_values > 0) {
      // TODO(wesm): Refactor to batch reads for improved memory use. It is not
      // trivial because the null_count is relative to the entire bitmap
      PARQUET_THROW_NOT_OK(indices_scratch_space_->TypedResize<int32_t>(
          num_values, /*shrink_to_fit=*/false));
    }

    auto indices_buffer =
        reinterpret_cast<int32_t*>(indices_scratch_space_->mutable_data());

    if (num_values != idx_decoder_.GetBatchSpaced(num_values, null_count, valid_bits,
                                                  valid_bits_offset, indices_buffer)) {
      ParquetException::EofException();
    }

    /// XXX(wesm): Cannot append "valid bits" directly to the builder
    std::vector<uint8_t> valid_bytes(num_values);
    ::arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values);
    for (int64_t i = 0; i < num_values; ++i) {
      valid_bytes[i] = static_cast<uint8_t>(bit_reader.IsSet());
      bit_reader.Next();
    }

    auto binary_builder = checked_cast<::arrow::BinaryDictionary32Builder*>(builder);
    PARQUET_THROW_NOT_OK(
        binary_builder->AppendIndices(indices_buffer, num_values, valid_bytes.data()));
    num_values_ -= num_values - null_count;
    return num_values - null_count;
  }

  int DecodeIndices(int num_values, ::arrow::ArrayBuilder* builder) override {
    num_values = std::min(num_values, num_values_);
    if (num_values > 0) {
      // TODO(wesm): Refactor to batch reads for improved memory use. This is
      // relatively simple here because we don't have to do any bookkeeping of
      // nulls
      PARQUET_THROW_NOT_OK(indices_scratch_space_->TypedResize<int32_t>(
          num_values, /*shrink_to_fit=*/false));
    }
    auto indices_buffer =
        reinterpret_cast<int32_t*>(indices_scratch_space_->mutable_data());
    if (num_values != idx_decoder_.GetBatch(indices_buffer, num_values)) {
      ParquetException::EofException();
    }
    auto binary_builder = checked_cast<::arrow::BinaryDictionary32Builder*>(builder);
    PARQUET_THROW_NOT_OK(binary_builder->AppendIndices(indices_buffer, num_values));
    num_values_ -= num_values;
    return num_values;
  }

  int DecodeIndices(int num_values, int32_t* indices) override {
    if (num_values != idx_decoder_.GetBatch(indices, num_values)) {
      ParquetException::EofException();
    }
    num_values_ -= num_values;
    return num_values;
  }

  void GetDictionary(const T** dictionary, int32_t* dictionary_length) override {
    *dictionary_length = dictionary_length_;
    *dictionary = reinterpret_cast<T*>(dictionary_->mutable_data());
  }

 protected:
  Status IndexInBounds(int32_t index) {
    if (ARROW_PREDICT_TRUE(0 <= index && index < dictionary_length_)) {
      return Status::OK();
    }
    return Status::Invalid("Index not in dictionary bounds");
  }

  inline void DecodeDict(TypedDecoder<Type>* dictionary) {
    dictionary_length_ = static_cast<int32_t>(dictionary->values_left());
    PARQUET_THROW_NOT_OK(dictionary_->Resize(dictionary_length_ * sizeof(T),
                                             /*shrink_to_fit=*/false));
    dictionary->Decode(reinterpret_cast<T*>(dictionary_->mutable_data()),
                       dictionary_length_);
  }

  // Only one is set.
  std::shared_ptr<ResizableBuffer> dictionary_;

  int32_t dictionary_length_;

  // Data that contains the byte array data (byte_array_dictionary_ just has the
  // pointers).
  std::shared_ptr<ResizableBuffer> byte_array_data_;

  // Arrow-style byte offsets for each dictionary value. We maintain two
  // representations of the dictionary, one as ByteArray* for non-Arrow
  // consumers and this one for Arrow consumers. Since dictionaries are
  // generally pretty small to begin with this doesn't mean too much extra
  // memory use in most cases
  std::shared_ptr<ResizableBuffer> byte_array_offsets_;

  // Reusable buffer for decoding dictionary indices to be appended to a
  // BinaryDictionary32Builder
  std::shared_ptr<ResizableBuffer> indices_scratch_space_;

  ::arrow::util::RleDecoder idx_decoder_;
};

template <typename Type>
void DictDecoderImpl<Type>::SetDict(TypedDecoder<Type>* dictionary) {
  DecodeDict(dictionary);
}

template <>
void DictDecoderImpl<BooleanType>::SetDict(TypedDecoder<BooleanType>* dictionary) {
  ParquetException::NYI("Dictionary encoding is not implemented for boolean values");
}

template <>
void DictDecoderImpl<ByteArrayType>::SetDict(TypedDecoder<ByteArrayType>* dictionary) {
  DecodeDict(dictionary);

  auto dict_values = reinterpret_cast<ByteArray*>(dictionary_->mutable_data());

  int total_size = 0;
  for (int i = 0; i < dictionary_length_; ++i) {
    total_size += dict_values[i].len;
  }
  PARQUET_THROW_NOT_OK(byte_array_data_->Resize(total_size,
                                                /*shrink_to_fit=*/false));
  PARQUET_THROW_NOT_OK(
      byte_array_offsets_->Resize((dictionary_length_ + 1) * sizeof(int32_t),
                                  /*shrink_to_fit=*/false));

  int32_t offset = 0;
  uint8_t* bytes_data = byte_array_data_->mutable_data();
  int32_t* bytes_offsets =
      reinterpret_cast<int32_t*>(byte_array_offsets_->mutable_data());
  for (int i = 0; i < dictionary_length_; ++i) {
    memcpy(bytes_data + offset, dict_values[i].ptr, dict_values[i].len);
    bytes_offsets[i] = offset;
    dict_values[i].ptr = bytes_data + offset;
    offset += dict_values[i].len;
  }
  bytes_offsets[dictionary_length_] = offset;
}

template <>
inline void DictDecoderImpl<FLBAType>::SetDict(TypedDecoder<FLBAType>* dictionary) {
  DecodeDict(dictionary);

  auto dict_values = reinterpret_cast<FLBA*>(dictionary_->mutable_data());

  int fixed_len = descr_->type_length();
  int total_size = dictionary_length_ * fixed_len;

  PARQUET_THROW_NOT_OK(byte_array_data_->Resize(total_size,
                                                /*shrink_to_fit=*/false));
  uint8_t* bytes_data = byte_array_data_->mutable_data();
  for (int32_t i = 0, offset = 0; i < dictionary_length_; ++i, offset += fixed_len) {
    memcpy(bytes_data + offset, dict_values[i].ptr, fixed_len);
    dict_values[i].ptr = bytes_data + offset;
  }
}

template <>
inline int DictDecoderImpl<Int96Type>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<Int96Type>::Accumulator* builder) {
  ParquetException::NYI("DecodeArrow to Int96Type");
}

template <>
inline int DictDecoderImpl<Int96Type>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<Int96Type>::DictAccumulator* builder) {
  ParquetException::NYI("DecodeArrow to Int96Type");
}

template <>
inline int DictDecoderImpl<ByteArrayType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<ByteArrayType>::Accumulator* builder) {
  ParquetException::NYI("DecodeArrow implemented elsewhere");
}

template <>
inline int DictDecoderImpl<ByteArrayType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<ByteArrayType>::DictAccumulator* builder) {
  ParquetException::NYI("DecodeArrow implemented elsewhere");
}

template <typename DType>
int DictDecoderImpl<DType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<DType>::DictAccumulator* builder) {
  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  auto dict_values = reinterpret_cast<const typename DType::c_type*>(dictionary_->data());

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        int32_t index;
        if (ARROW_PREDICT_FALSE(!idx_decoder_.Get(&index))) {
          throw ParquetException("");
        }
        PARQUET_THROW_NOT_OK(IndexInBounds(index));
        PARQUET_THROW_NOT_OK(builder->Append(dict_values[index]));
      },
      [&]() { PARQUET_THROW_NOT_OK(builder->AppendNull()); });

  return num_values - null_count;
}

template <>
int DictDecoderImpl<BooleanType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<BooleanType>::DictAccumulator* builder) {
  ParquetException::NYI("No dictionary encoding for BooleanType");
}

template <>
inline int DictDecoderImpl<FLBAType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<FLBAType>::Accumulator* builder) {
  if (builder->byte_width() != descr_->type_length()) {
    throw ParquetException("Byte width mismatch: builder was " +
                           std::to_string(builder->byte_width()) + " but decoder was " +
                           std::to_string(descr_->type_length()));
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  auto dict_values = reinterpret_cast<const FLBA*>(dictionary_->data());

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        int32_t index;
        if (ARROW_PREDICT_FALSE(!idx_decoder_.Get(&index))) {
          throw ParquetException("");
        }
        PARQUET_THROW_NOT_OK(IndexInBounds(index));
        builder->UnsafeAppend(dict_values[index].ptr);
      },
      [&]() { builder->UnsafeAppendNull(); });

  return num_values - null_count;
}

template <>
int DictDecoderImpl<FLBAType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<FLBAType>::DictAccumulator* builder) {
  auto value_type =
      checked_cast<const ::arrow::DictionaryType&>(*builder->type()).value_type();
  auto byte_width =
      checked_cast<const ::arrow::FixedSizeBinaryType&>(*value_type).byte_width();
  if (byte_width != descr_->type_length()) {
    throw ParquetException("Byte width mismatch: builder was " +
                           std::to_string(byte_width) + " but decoder was " +
                           std::to_string(descr_->type_length()));
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  auto dict_values = reinterpret_cast<const FLBA*>(dictionary_->data());

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        int32_t index;
        if (ARROW_PREDICT_FALSE(!idx_decoder_.Get(&index))) {
          throw ParquetException("");
        }
        PARQUET_THROW_NOT_OK(IndexInBounds(index));
        PARQUET_THROW_NOT_OK(builder->Append(dict_values[index].ptr));
      },
      [&]() { PARQUET_THROW_NOT_OK(builder->AppendNull()); });

  return num_values - null_count;
}

template <typename Type>
int DictDecoderImpl<Type>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<Type>::Accumulator* builder) {
  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  using value_type = typename Type::c_type;
  auto dict_values = reinterpret_cast<const value_type*>(dictionary_->data());

  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        int32_t index;
        if (ARROW_PREDICT_FALSE(!idx_decoder_.Get(&index))) {
          throw ParquetException("");
        }
        PARQUET_THROW_NOT_OK(IndexInBounds(index));
        builder->UnsafeAppend(dict_values[index]);
      },
      [&]() { builder->UnsafeAppendNull(); });

  return num_values - null_count;
}

template <typename Type>
void DictDecoderImpl<Type>::InsertDictionary(::arrow::ArrayBuilder* builder) {
  ParquetException::NYI("InsertDictionary only implemented for BYTE_ARRAY types");
}

template <>
void DictDecoderImpl<ByteArrayType>::InsertDictionary(::arrow::ArrayBuilder* builder) {
  auto binary_builder = checked_cast<::arrow::BinaryDictionary32Builder*>(builder);

  // Make a BinaryArray referencing the internal dictionary data
  auto arr = std::make_shared<::arrow::BinaryArray>(
      dictionary_length_, byte_array_offsets_, byte_array_data_);
  PARQUET_THROW_NOT_OK(binary_builder->InsertMemoValues(*arr));
}

class DictByteArrayDecoderImpl : public DictDecoderImpl<ByteArrayType>,
                                 virtual public ByteArrayDecoder {
 public:
  using BASE = DictDecoderImpl<ByteArrayType>;
  using BASE::DictDecoderImpl;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  ::arrow::BinaryDictionary32Builder* builder) override {
    int result = 0;
    if (null_count == 0) {
      PARQUET_THROW_NOT_OK(DecodeArrowNonNull(num_values, builder, &result));
    } else {
      PARQUET_THROW_NOT_OK(DecodeArrow(num_values, null_count, valid_bits,
                                       valid_bits_offset, builder, &result));
    }
    return result;
  }

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<ByteArrayType>::Accumulator* out) override {
    int result = 0;
    if (null_count == 0) {
      PARQUET_THROW_NOT_OK(DecodeArrowDenseNonNull(num_values, out, &result));
    } else {
      PARQUET_THROW_NOT_OK(DecodeArrowDense(num_values, null_count, valid_bits,
                                            valid_bits_offset, out, &result));
    }
    return result;
  }

 private:
  Status DecodeArrowDense(int num_values, int null_count, const uint8_t* valid_bits,
                          int64_t valid_bits_offset,
                          typename EncodingTraits<ByteArrayType>::Accumulator* out,
                          int* out_num_values) {
    constexpr int32_t kBufferSize = 1024;
    int32_t indices[kBufferSize];

    ArrowBinaryHelper helper(out);

    ::arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values);

    auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data());
    int values_decoded = 0;
    int num_appended = 0;
    while (num_appended < num_values) {
      bool is_valid = bit_reader.IsSet();
      bit_reader.Next();

      if (is_valid) {
        int32_t batch_size =
            std::min<int32_t>(kBufferSize, num_values - num_appended - null_count);
        int num_indices = idx_decoder_.GetBatch(indices, batch_size);

        if (ARROW_PREDICT_FALSE(num_indices < 1)) {
          return Status::Invalid("Invalid number of indices '", num_indices, "'");
        }

        int i = 0;
        while (true) {
          // Consume all indices
          if (is_valid) {
            auto idx = indices[i];
            RETURN_NOT_OK(IndexInBounds(idx));
            const auto& val = dict_values[idx];
            if (ARROW_PREDICT_FALSE(!helper.CanFit(val.len))) {
              RETURN_NOT_OK(helper.PushChunk());
            }
            RETURN_NOT_OK(helper.Append(val.ptr, static_cast<int32_t>(val.len)));
            ++i;
            ++values_decoded;
          } else {
            RETURN_NOT_OK(helper.AppendNull());
            --null_count;
          }
          ++num_appended;
          if (i == num_indices) {
            // Do not advance the bit_reader if we have fulfilled the decode
            // request
            break;
          }
          is_valid = bit_reader.IsSet();
          bit_reader.Next();
        }
      } else {
        RETURN_NOT_OK(helper.AppendNull());
        --null_count;
        ++num_appended;
      }
    }
    *out_num_values = values_decoded;
    return Status::OK();
  }

  Status DecodeArrowDenseNonNull(int num_values,
                                 typename EncodingTraits<ByteArrayType>::Accumulator* out,
                                 int* out_num_values) {
    constexpr int32_t kBufferSize = 2048;
    int32_t indices[kBufferSize];
    int values_decoded = 0;

    ArrowBinaryHelper helper(out);
    auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data());

    while (values_decoded < num_values) {
      int32_t batch_size = std::min<int32_t>(kBufferSize, num_values - values_decoded);
      int num_indices = idx_decoder_.GetBatch(indices, batch_size);
      if (num_indices == 0) ParquetException::EofException();
      for (int i = 0; i < num_indices; ++i) {
        auto idx = indices[i];
        RETURN_NOT_OK(IndexInBounds(idx));
        const auto& val = dict_values[idx];
        if (ARROW_PREDICT_FALSE(!helper.CanFit(val.len))) {
          RETURN_NOT_OK(helper.PushChunk());
        }
        RETURN_NOT_OK(helper.Append(val.ptr, static_cast<int32_t>(val.len)));
      }
      values_decoded += num_indices;
    }
    *out_num_values = values_decoded;
    return Status::OK();
  }

  template <typename BuilderType>
  Status DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                     int64_t valid_bits_offset, BuilderType* builder,
                     int* out_num_values) {
    constexpr int32_t kBufferSize = 1024;
    int32_t indices[kBufferSize];

    RETURN_NOT_OK(builder->Reserve(num_values));
    ::arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values);

    auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data());

    int values_decoded = 0;
    int num_appended = 0;
    while (num_appended < num_values) {
      bool is_valid = bit_reader.IsSet();
      bit_reader.Next();

      if (is_valid) {
        int32_t batch_size =
            std::min<int32_t>(kBufferSize, num_values - num_appended - null_count);
        int num_indices = idx_decoder_.GetBatch(indices, batch_size);

        int i = 0;
        while (true) {
          // Consume all indices
          if (is_valid) {
            auto idx = indices[i];
            RETURN_NOT_OK(IndexInBounds(idx));
            const auto& val = dict_values[idx];
            RETURN_NOT_OK(builder->Append(val.ptr, val.len));
            ++i;
            ++values_decoded;
          } else {
            RETURN_NOT_OK(builder->AppendNull());
            --null_count;
          }
          ++num_appended;
          if (i == num_indices) {
            // Do not advance the bit_reader if we have fulfilled the decode
            // request
            break;
          }
          is_valid = bit_reader.IsSet();
          bit_reader.Next();
        }
      } else {
        RETURN_NOT_OK(builder->AppendNull());
        --null_count;
        ++num_appended;
      }
    }
    *out_num_values = values_decoded;
    return Status::OK();
  }

  template <typename BuilderType>
  Status DecodeArrowNonNull(int num_values, BuilderType* builder, int* out_num_values) {
    constexpr int32_t kBufferSize = 2048;
    int32_t indices[kBufferSize];

    RETURN_NOT_OK(builder->Reserve(num_values));

    auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data());

    int values_decoded = 0;
    while (values_decoded < num_values) {
      int32_t batch_size = std::min<int32_t>(kBufferSize, num_values - values_decoded);
      int num_indices = idx_decoder_.GetBatch(indices, batch_size);
      if (num_indices == 0) ParquetException::EofException();
      for (int i = 0; i < num_indices; ++i) {
        auto idx = indices[i];
        RETURN_NOT_OK(IndexInBounds(idx));
        const auto& val = dict_values[idx];
        RETURN_NOT_OK(builder->Append(val.ptr, val.len));
      }
      values_decoded += num_indices;
    }
    *out_num_values = values_decoded;
    return Status::OK();
  }
};

// ----------------------------------------------------------------------
// DeltaBitPackDecoder

template <typename DType>
class DeltaBitPackDecoder : public DecoderImpl, virtual public TypedDecoder<DType> {
 public:
  typedef typename DType::c_type T;

  explicit DeltaBitPackDecoder(const ColumnDescriptor* descr,
                               MemoryPool* pool = ::arrow::default_memory_pool())
      : DecoderImpl(descr, Encoding::DELTA_BINARY_PACKED), pool_(pool) {
    if (DType::type_num != Type::INT32 && DType::type_num != Type::INT64) {
      throw ParquetException("Delta bit pack encoding should only be for integer data.");
    }
  }

  void SetData(int num_values, const uint8_t* data, int len) override {
    this->num_values_ = num_values;
    decoder_ = ::arrow::BitUtil::BitReader(data, len);
    values_current_block_ = 0;
    values_current_mini_block_ = 0;
  }

  int Decode(T* buffer, int max_values) override {
    return GetInternal(buffer, max_values);
  }

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<DType>::Accumulator* out) override {
    if (null_count != 0) {
      ParquetException::NYI("Delta bit pack DecodeArrow with null slots");
    }
    std::vector<T> values(num_values);
    GetInternal(values.data(), num_values);
    PARQUET_THROW_NOT_OK(out->AppendValues(values));
    return num_values;
  }

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<DType>::DictAccumulator* out) override {
    if (null_count != 0) {
      ParquetException::NYI("Delta bit pack DecodeArrow with null slots");
    }
    std::vector<T> values(num_values);
    GetInternal(values.data(), num_values);
    PARQUET_THROW_NOT_OK(out->Reserve(num_values));
    for (T value : values) {
      PARQUET_THROW_NOT_OK(out->Append(value));
    }
    return num_values;
  }

 private:
  void InitBlock() {
    // The number of values per block.
    uint32_t block_size;
    if (!decoder_.GetVlqInt(&block_size)) ParquetException::EofException();
    if (!decoder_.GetVlqInt(&num_mini_blocks_)) ParquetException::EofException();
    if (!decoder_.GetVlqInt(&values_current_block_)) {
      ParquetException::EofException();
    }
    if (!decoder_.GetZigZagVlqInt(&last_value_)) ParquetException::EofException();

    delta_bit_widths_ = AllocateBuffer(pool_, num_mini_blocks_);
    uint8_t* bit_width_data = delta_bit_widths_->mutable_data();

    if (!decoder_.GetZigZagVlqInt(&min_delta_)) ParquetException::EofException();
    for (uint32_t i = 0; i < num_mini_blocks_; ++i) {
      if (!decoder_.GetAligned<uint8_t>(1, bit_width_data + i)) {
        ParquetException::EofException();
      }
    }
    values_per_mini_block_ = block_size / num_mini_blocks_;
    mini_block_idx_ = 0;
    delta_bit_width_ = bit_width_data[0];
    values_current_mini_block_ = values_per_mini_block_;
  }

  template <typename T>
  int GetInternal(T* buffer, int max_values) {
    max_values = std::min(max_values, this->num_values_);
    const uint8_t* bit_width_data = delta_bit_widths_->data();
    for (int i = 0; i < max_values; ++i) {
      if (ARROW_PREDICT_FALSE(values_current_mini_block_ == 0)) {
        ++mini_block_idx_;
        if (mini_block_idx_ < static_cast<size_t>(delta_bit_widths_->size())) {
          delta_bit_width_ = bit_width_data[mini_block_idx_];
          values_current_mini_block_ = values_per_mini_block_;
        } else {
          InitBlock();
          buffer[i] = last_value_;
          continue;
        }
      }

      // TODO: the key to this algorithm is to decode the entire miniblock at once.
      int64_t delta;
      if (!decoder_.GetValue(delta_bit_width_, &delta)) ParquetException::EofException();
      delta += min_delta_;
      last_value_ += static_cast<int32_t>(delta);
      buffer[i] = last_value_;
      --values_current_mini_block_;
    }
    this->num_values_ -= max_values;
    return max_values;
  }

  MemoryPool* pool_;
  ::arrow::BitUtil::BitReader decoder_;
  uint32_t values_current_block_;
  uint32_t num_mini_blocks_;
  uint64_t values_per_mini_block_;
  uint64_t values_current_mini_block_;

  int32_t min_delta_;
  size_t mini_block_idx_;
  std::shared_ptr<ResizableBuffer> delta_bit_widths_;
  int delta_bit_width_;

  int32_t last_value_;
};

// ----------------------------------------------------------------------
// DELTA_LENGTH_BYTE_ARRAY

class DeltaLengthByteArrayDecoder : public DecoderImpl,
                                    virtual public TypedDecoder<ByteArrayType> {
 public:
  explicit DeltaLengthByteArrayDecoder(const ColumnDescriptor* descr,
                                       MemoryPool* pool = ::arrow::default_memory_pool())
      : DecoderImpl(descr, Encoding::DELTA_LENGTH_BYTE_ARRAY),
        len_decoder_(nullptr, pool),
        pool_(pool) {}

  void SetData(int num_values, const uint8_t* data, int len) override {
    num_values_ = num_values;
    if (len == 0) return;
    int total_lengths_len = ::arrow::util::SafeLoadAs<int32_t>(data);
    data += 4;
    this->len_decoder_.SetData(num_values, data, total_lengths_len);
    data_ = data + total_lengths_len;
    this->len_ = len - 4 - total_lengths_len;
  }

  int Decode(ByteArray* buffer, int max_values) override {
    using VectorT = ArrowPoolVector<int>;
    max_values = std::min(max_values, num_values_);
    VectorT lengths(max_values, 0, ::arrow::stl::allocator<int>(pool_));
    len_decoder_.Decode(lengths.data(), max_values);
    for (int i = 0; i < max_values; ++i) {
      buffer[i].len = lengths[i];
      buffer[i].ptr = data_;
      this->data_ += lengths[i];
      this->len_ -= lengths[i];
    }
    this->num_values_ -= max_values;
    return max_values;
  }

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<ByteArrayType>::Accumulator* out) override {
    ParquetException::NYI("DecodeArrow for DeltaLengthByteArrayDecoder");
  }

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<ByteArrayType>::DictAccumulator* out) override {
    ParquetException::NYI("DecodeArrow for DeltaLengthByteArrayDecoder");
  }

 private:
  DeltaBitPackDecoder<Int32Type> len_decoder_;
  ::arrow::MemoryPool* pool_;
};

// ----------------------------------------------------------------------
// DELTA_BYTE_ARRAY

class DeltaByteArrayDecoder : public DecoderImpl,
                              virtual public TypedDecoder<ByteArrayType> {
 public:
  explicit DeltaByteArrayDecoder(const ColumnDescriptor* descr,
                                 MemoryPool* pool = ::arrow::default_memory_pool())
      : DecoderImpl(descr, Encoding::DELTA_BYTE_ARRAY),
        prefix_len_decoder_(nullptr, pool),
        suffix_decoder_(nullptr, pool),
        last_value_(0, nullptr) {}

  virtual void SetData(int num_values, const uint8_t* data, int len) {
    num_values_ = num_values;
    if (len == 0) return;
    int prefix_len_length = ::arrow::util::SafeLoadAs<int32_t>(data);
    data += 4;
    len -= 4;
    prefix_len_decoder_.SetData(num_values, data, prefix_len_length);
    data += prefix_len_length;
    len -= prefix_len_length;
    suffix_decoder_.SetData(num_values, data, len);
  }

  // TODO: this doesn't work and requires memory management. We need to allocate
  // new strings to store the results.
  virtual int Decode(ByteArray* buffer, int max_values) {
    max_values = std::min(max_values, this->num_values_);
    for (int i = 0; i < max_values; ++i) {
      int prefix_len = 0;
      prefix_len_decoder_.Decode(&prefix_len, 1);
      ByteArray suffix = {0, nullptr};
      suffix_decoder_.Decode(&suffix, 1);
      buffer[i].len = prefix_len + suffix.len;

      uint8_t* result = reinterpret_cast<uint8_t*>(malloc(buffer[i].len));
      memcpy(result, last_value_.ptr, prefix_len);
      memcpy(result + prefix_len, suffix.ptr, suffix.len);

      buffer[i].ptr = result;
      last_value_ = buffer[i];
    }
    this->num_values_ -= max_values;
    return max_values;
  }

 private:
  DeltaBitPackDecoder<Int32Type> prefix_len_decoder_;
  DeltaLengthByteArrayDecoder suffix_decoder_;
  ByteArray last_value_;
};

// ----------------------------------------------------------------------
// BYTE_STREAM_SPLIT

template <typename DType>
class ByteStreamSplitDecoder : public DecoderImpl, virtual public TypedDecoder<DType> {
 public:
  using T = typename DType::c_type;
  explicit ByteStreamSplitDecoder(const ColumnDescriptor* descr);

  int Decode(T* buffer, int max_values) override;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<DType>::Accumulator* builder) override;

  int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits,
                  int64_t valid_bits_offset,
                  typename EncodingTraits<DType>::DictAccumulator* builder) override;

  void SetData(int num_values, const uint8_t* data, int len) override;

  T* EnsureDecodeBuffer(int64_t min_values) {
    const int64_t size = sizeof(T) * min_values;
    if (!decode_buffer_ || decode_buffer_->size() < size) {
      PARQUET_ASSIGN_OR_THROW(decode_buffer_, ::arrow::AllocateBuffer(size));
    }
    return reinterpret_cast<T*>(decode_buffer_->mutable_data());
  }

 private:
  int num_values_in_buffer_{0};
  std::shared_ptr<Buffer> decode_buffer_;

  static constexpr size_t kNumStreams = sizeof(T);
};

template <typename DType>
ByteStreamSplitDecoder<DType>::ByteStreamSplitDecoder(const ColumnDescriptor* descr)
    : DecoderImpl(descr, Encoding::BYTE_STREAM_SPLIT) {}

template <typename DType>
void ByteStreamSplitDecoder<DType>::SetData(int num_values, const uint8_t* data,
                                            int len) {
  DecoderImpl::SetData(num_values, data, len);
  if (num_values * static_cast<int64_t>(sizeof(T)) > len) {
    throw ParquetException("Data size too small for number of values (corrupted file?)");
  }
  num_values_in_buffer_ = num_values;
}

template <typename DType>
int ByteStreamSplitDecoder<DType>::Decode(T* buffer, int max_values) {
  const int values_to_decode = std::min(num_values_, max_values);
  const int num_decoded_previously = num_values_in_buffer_ - num_values_;
  const uint8_t* data = data_ + num_decoded_previously;

  ::arrow::util::internal::ByteStreamSplitDecode<T>(data, values_to_decode,
                                                    num_values_in_buffer_, buffer);
  num_values_ -= values_to_decode;
  len_ -= sizeof(T) * values_to_decode;
  return values_to_decode;
}

template <typename DType>
int ByteStreamSplitDecoder<DType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<DType>::Accumulator* builder) {
  constexpr int value_size = static_cast<int>(kNumStreams);
  int values_decoded = num_values - null_count;
  if (ARROW_PREDICT_FALSE(len_ < value_size * values_decoded)) {
    ParquetException::EofException();
  }

  PARQUET_THROW_NOT_OK(builder->Reserve(num_values));

  const int num_decoded_previously = num_values_in_buffer_ - num_values_;
  const uint8_t* data = data_ + num_decoded_previously;
  int offset = 0;

#if defined(ARROW_HAVE_SIMD_SPLIT)
  // Use fast decoding into intermediate buffer.  This will also decode
  // some null values, but it's fast enough that we don't care.
  T* decode_out = EnsureDecodeBuffer(values_decoded);
  ::arrow::util::internal::ByteStreamSplitDecode<T>(data, values_decoded,
                                                    num_values_in_buffer_, decode_out);

  // XXX If null_count is 0, we could even append in bulk or decode directly into
  // builder
  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        builder->UnsafeAppend(decode_out[offset]);
        ++offset;
      },
      [&]() { builder->UnsafeAppendNull(); });

#else
  VisitNullBitmapInline(
      valid_bits, valid_bits_offset, num_values, null_count,
      [&]() {
        uint8_t gathered_byte_data[kNumStreams];
        for (size_t b = 0; b < kNumStreams; ++b) {
          const size_t byte_index = b * num_values_in_buffer_ + offset;
          gathered_byte_data[b] = data[byte_index];
        }
        builder->UnsafeAppend(::arrow::util::SafeLoadAs<T>(&gathered_byte_data[0]));
        ++offset;
      },
      [&]() { builder->UnsafeAppendNull(); });
#endif

  num_values_ -= values_decoded;
  len_ -= sizeof(T) * values_decoded;
  return values_decoded;
}

template <typename DType>
int ByteStreamSplitDecoder<DType>::DecodeArrow(
    int num_values, int null_count, const uint8_t* valid_bits, int64_t valid_bits_offset,
    typename EncodingTraits<DType>::DictAccumulator* builder) {
  ParquetException::NYI("DecodeArrow for ByteStreamSplitDecoder");
}

}  // namespace

// ----------------------------------------------------------------------
// Encoder and decoder factory functions

std::unique_ptr<Encoder> MakeEncoder(Type::type type_num, Encoding::type encoding,
                                     bool use_dictionary, const ColumnDescriptor* descr,
                                     MemoryPool* pool) {
  if (use_dictionary) {
    switch (type_num) {
      case Type::INT32:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<Int32Type>(descr, pool));
      case Type::INT64:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<Int64Type>(descr, pool));
      case Type::INT96:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<Int96Type>(descr, pool));
      case Type::FLOAT:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<FloatType>(descr, pool));
      case Type::DOUBLE:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<DoubleType>(descr, pool));
      case Type::BYTE_ARRAY:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<ByteArrayType>(descr, pool));
      case Type::FIXED_LEN_BYTE_ARRAY:
        return std::unique_ptr<Encoder>(new DictEncoderImpl<FLBAType>(descr, pool));
      default:
        DCHECK(false) << "Encoder not implemented";
        break;
    }
  } else if (encoding == Encoding::PLAIN) {
    switch (type_num) {
      case Type::BOOLEAN:
        return std::unique_ptr<Encoder>(new PlainEncoder<BooleanType>(descr, pool));
      case Type::INT32:
        return std::unique_ptr<Encoder>(new PlainEncoder<Int32Type>(descr, pool));
      case Type::INT64:
        return std::unique_ptr<Encoder>(new PlainEncoder<Int64Type>(descr, pool));
      case Type::INT96:
        return std::unique_ptr<Encoder>(new PlainEncoder<Int96Type>(descr, pool));
      case Type::FLOAT:
        return std::unique_ptr<Encoder>(new PlainEncoder<FloatType>(descr, pool));
      case Type::DOUBLE:
        return std::unique_ptr<Encoder>(new PlainEncoder<DoubleType>(descr, pool));
      case Type::BYTE_ARRAY:
        return std::unique_ptr<Encoder>(new PlainEncoder<ByteArrayType>(descr, pool));
      case Type::FIXED_LEN_BYTE_ARRAY:
        return std::unique_ptr<Encoder>(new PlainEncoder<FLBAType>(descr, pool));
      default:
        DCHECK(false) << "Encoder not implemented";
        break;
    }
  } else if (encoding == Encoding::BYTE_STREAM_SPLIT) {
    switch (type_num) {
      case Type::FLOAT:
        return std::unique_ptr<Encoder>(
            new ByteStreamSplitEncoder<FloatType>(descr, pool));
      case Type::DOUBLE:
        return std::unique_ptr<Encoder>(
            new ByteStreamSplitEncoder<DoubleType>(descr, pool));
      default:
        throw ParquetException("BYTE_STREAM_SPLIT only supports FLOAT and DOUBLE");
        break;
    }
  } else {
    ParquetException::NYI("Selected encoding is not supported");
  }
  DCHECK(false) << "Should not be able to reach this code";
  return nullptr;
}

std::unique_ptr<Decoder> MakeDecoder(Type::type type_num, Encoding::type encoding,
                                     const ColumnDescriptor* descr) {
  if (encoding == Encoding::PLAIN) {
    switch (type_num) {
      case Type::BOOLEAN:
        return std::unique_ptr<Decoder>(new PlainBooleanDecoder(descr));
      case Type::INT32:
        return std::unique_ptr<Decoder>(new PlainDecoder<Int32Type>(descr));
      case Type::INT64:
        return std::unique_ptr<Decoder>(new PlainDecoder<Int64Type>(descr));
      case Type::INT96:
        return std::unique_ptr<Decoder>(new PlainDecoder<Int96Type>(descr));
      case Type::FLOAT:
        return std::unique_ptr<Decoder>(new PlainDecoder<FloatType>(descr));
      case Type::DOUBLE:
        return std::unique_ptr<Decoder>(new PlainDecoder<DoubleType>(descr));
      case Type::BYTE_ARRAY:
        return std::unique_ptr<Decoder>(new PlainByteArrayDecoder(descr));
      case Type::FIXED_LEN_BYTE_ARRAY:
        return std::unique_ptr<Decoder>(new PlainFLBADecoder(descr));
      default:
        break;
    }
  } else if (encoding == Encoding::BYTE_STREAM_SPLIT) {
    switch (type_num) {
      case Type::FLOAT:
        return std::unique_ptr<Decoder>(new ByteStreamSplitDecoder<FloatType>(descr));
      case Type::DOUBLE:
        return std::unique_ptr<Decoder>(new ByteStreamSplitDecoder<DoubleType>(descr));
      default:
        throw ParquetException("BYTE_STREAM_SPLIT only supports FLOAT and DOUBLE");
        break;
    }
  } else {
    ParquetException::NYI("Selected encoding is not supported");
  }
  DCHECK(false) << "Should not be able to reach this code";
  return nullptr;
}

namespace detail {
std::unique_ptr<Decoder> MakeDictDecoder(Type::type type_num,
                                         const ColumnDescriptor* descr,
                                         MemoryPool* pool) {
  switch (type_num) {
    case Type::BOOLEAN:
      ParquetException::NYI("Dictionary encoding not implemented for boolean type");
    case Type::INT32:
      return std::unique_ptr<Decoder>(new DictDecoderImpl<Int32Type>(descr, pool));
    case Type::INT64:
      return std::unique_ptr<Decoder>(new DictDecoderImpl<Int64Type>(descr, pool));
    case Type::INT96:
      return std::unique_ptr<Decoder>(new DictDecoderImpl<Int96Type>(descr, pool));
    case Type::FLOAT:
      return std::unique_ptr<Decoder>(new DictDecoderImpl<FloatType>(descr, pool));
    case Type::DOUBLE:
      return std::unique_ptr<Decoder>(new DictDecoderImpl<DoubleType>(descr, pool));
    case Type::BYTE_ARRAY:
      return std::unique_ptr<Decoder>(new DictByteArrayDecoderImpl(descr, pool));
    case Type::FIXED_LEN_BYTE_ARRAY:
      return std::unique_ptr<Decoder>(new DictDecoderImpl<FLBAType>(descr, pool));
    default:
      break;
  }
  DCHECK(false) << "Should not be able to reach this code";
  return nullptr;
}

}  // namespace detail
}  // namespace parquet