| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch adds a support for a basic aggregation on `YqlSelect` and
includes translation, type annotation, expansion changes.
There are `count`, `min`, `max`, `sum`, `avg` and other aggregations
without parameters supported.
`HAVING` support will be added in a future PRs. Support for (2+)-arg
aggregation is postponed, as it is not required for TPCH queries.
### Translator
The `std::expected` was passed though the function call hierarchy to
gracefully do not support some aggregation functions.
The translation unit `select_yql_aggregation` was introduced.
It implements an alternative aggregation translation to
`YqlAggFactory` and `YqlAgg` callables.
It reuses an `extractor` body and `aggregation_traits_factory` from
a legacy code, via dirty hack with a `friend` keyword, as I decided
to make minimal changes on an existing code.
For a query fragment:
```yql
... Sum(body) ...
```
Resulting YQLs AST looks like this:
```yqls
(YqlAggApply
(YqlAggFactory '"sum")
'() # <-- Options
(Void) # <-- Result type stub
body')
```
### Type Annotation
I followed a plan described at the https://nda.ya.ru/t/psA5Gaji7PDe9R.
Besides, new `PgXXX` to `YqlXXX` renames, there were introduced
`YqlAggFactoryWrapper` and `YqlAggWrapper`. First just checks
argument count and types, but the second does more sofisticated work.
`YqlAggWrapper` acts in 2 stages. On the first stage it will write
an expression with a `TypeOf` of a `Result` to a result type stub
and then calls for a transformation pipeline repetition.
On the second stage result type stub is typed and so it
just defines its type to be equal to a type of the stub.
The tricky thing here is to contruct this expression with `Result` type.
To do this it needs to instantiate an aggregation traits factory with a
correct `list_type` and an `extractor`. The `list_type` is just a
`(ListType (TypeOf row))`, where `row` is an `Argument` from an enclosing
`lambda` at `YqlResultItem`. An `extractor` is just a `(lambda (row) body')`.
The traits are imported from the `mount/lib/aggregate.yqls`, then they are
beta-reduced with constructed `list_type` and `extractor` expressions. Then
the `state` method is extracted and beta-reduced with a `body`. Then the
`finish` method is extracted and beta-reduced with a some `state`-expression.
So the resulting expression is constructed.
### Expansion
The `YqlAgg` is expanded at `BuildAggregationTraits` from `yql_co_pgselect`,
that is nicely integrated into `PgSelect` expansion infrastructure. There it
is sufficient just to import an aggregation traits factory and beta-reduce it
with `list_type` and `extractor`, that are provided by the `PgSelect`-related
code.
### Refactoring
Loading an `ExprNode`s from a module, so I extracted a logic to
`ImportReadonly` and `ImportDeeplyCopied` at `yql_module_helpers`.
commit_hash:9303f00567dd423bedc334e7c7514584f5bd8cff
|
