spark-instrumented-optimizer/sql/core
Chao Sun 86d3bb5f7d [SPARK-34981][SQL] Implement V2 function resolution and evaluation
Co-Authored-By: Chao Sun <sunchaoapple.com>
Co-Authored-By: Ryan Blue <rbluenetflix.com>

### What changes were proposed in this pull request?

This implements function resolution and evaluation for functions registered through V2 FunctionCatalog [SPARK-27658](https://issues.apache.org/jira/browse/SPARK-27658). In particular:
- Added documentation for how to define the "magic method" in `ScalarFunction`.
- Added a new expression `ApplyFunctionExpression` which evaluates input by delegating to `ScalarFunction.produceResult` method.
- added a new expression `V2Aggregator` which is a type of `TypedImperativeAggregate`. It's a wrapper of V2 `AggregateFunction` and mostly delegate methods to the implementation of the latter. It also uses plain Java serde for intermediate state.
- Added function resolution logic for `ScalarFunction` and `AggregateFunction` in `Analyzer`.
  + For `ScalarFunction` this checks if the magic method is implemented through Java reflection, and create a `Invoke` expression if so. Otherwise, it checks if the default `produceResult` is overridden. If so, it creates a `ApplyFunctionExpression` which evaluates through `InternalRow`. Otherwise an analysis exception is thrown.
 + For `AggregateFunction`, this checks if the `update` method is overridden. If so, it converts it to `V2Aggregator`. Otherwise an analysis exception is thrown similar to the case of `ScalarFunction`.
- Extended existing `InMemoryTableCatalog` to add the function catalog capability. Also renamed it to `InMemoryCatalog` since it no longer only covers tables.

**Note**: this currently can successfully detect whether a subclass overrides the default `produceResult` or `update` method from the parent interface **only for Java implementations**. It seems in Scala it's hard to differentiate whether a subclass overrides a default method from its parent interface. In this case, it will be a runtime error instead of analysis error.

A few TODOs:
- Extend `V2SessionCatalog` with function catalog. This seems a little tricky since API such V2 `FunctionCatalog`'s `loadFunction` is different from V1 `SessionCatalog`'s `lookupFunction`.
- Add magic method for `AggregateFunction`.
- Type coercion when looking up functions

### Why are the changes needed?

As V2 FunctionCatalog APIs are finalized, we should integrate it with function resolution and evaluation process so that they are actually useful.

### Does this PR introduce _any_ user-facing change?

Yes, now a function exposed through V2 FunctionCatalog can be analyzed and evaluated.

### How was this patch tested?

Added new unit tests.

Closes #32082 from sunchao/resolve-func-v2.

Lead-authored-by: Chao Sun <sunchao@apple.com>
Co-authored-by: Chao Sun <sunchao@apache.org>
Co-authored-by: Chao Sun <sunchao@uber.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
2021-04-28 17:21:49 +00:00
..
benchmarks [SPARK-35235][SQL][TEST] Add row-based hash map into aggregate benchmark 2021-04-27 06:53:42 +00:00
src [SPARK-34981][SQL] Implement V2 function resolution and evaluation 2021-04-28 17:21:49 +00:00
pom.xml [SPARK-33662][BUILD] Setting version to 3.2.0-SNAPSHOT 2020-12-04 14:10:42 -08:00