pedrumj2 opened a new pull request, #13603: URL: https://github.com/apache/gluten/pull/13603
## What changes are proposed in this pull request? Native [filter](https://spark.apache.org/docs/3.5.6/api/sql/index.html#filter) now returns Spark's result when its lambda applies `IS NOT NULL` to a struct field, an outer column, or anything other than the element itself; only `filter(arr, x -> x IS NOT NULL)` still runs as [array_compact](https://spark.apache.org/docs/3.5.6/api/sql/index.html#array_compact). apache/gluten#9349 made [genArrayFilterTransformer](https://github.com/apache/gluten/blob/cdf5292884895126825e4349938f2da3959fc812/backends-velox/src/main/scala/org/apache/gluten/backendsapi/velox/VeloxSparkPlanExecApi.scala#L215-L218) run a `filter` as `array_compact` whenever the lambda body is [IsNotNull](https://github.com/apache/spark/blob/v3.5.5/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/nullExpressions.scala#L380), whatever that `IsNotNull` tests. [Spark's array_compact](https://github.com/apache/spark/blob/v3.5.5/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/collectionOperations.scala#L5050-L5052) removes null elements, and Spark defines it as a `filter` whose lambda tests the element itself. So these two queries return the same rows, and only this `filter` can run as `array_compact`: ```sql SELECT array_compact(arr) FROM t; SELECT filter(arr, x -> x IS NOT NULL) FROM t; ``` A filter that tests a struct field or another column also runs as `array_compact`, and returns the wrong rows. For an `arr` of `[{"a":1},{"a":null}]` and a `c` of `NULL`: | Query | Spark | Gluten before this PR | | --- | --- | --- | | `filter(arr, x -> x IS NOT NULL)` | `[{"a":1},{"a":null}]` | `[{"a":1},{"a":null}]` | | `filter(arr, x -> x.a IS NOT NULL)` | `[{"a":1}]` | `[{"a":1},{"a":null}]` | | `filter(arr, x -> c IS NOT NULL)` | `[]` | `[{"a":1},{"a":null}]` | This PR runs a `filter` as `array_compact` only when its `IS NOT NULL` tests the lambda variable itself, in the Velox backend and in the same rewrite in [BoltSparkPlanExecApi](https://github.com/apache/gluten/blob/cdf5292884895126825e4349938f2da3959fc812/backends-bolt/src/main/scala/org/apache/gluten/backendsapi/bolt/BoltSparkPlanExecApi.scala#L194). The second and third queries then run as `filter` and return Spark's rows, still natively. ## How was this patch tested? ### Reproduction 1. Clone the repo and check out the base commit: ```bash git clone https://github.com/apache/gluten.git cd gluten git checkout cdf5292884895126825e4349938f2da3959fc812 ``` 2. Build the native libraries and the bundle jar: ```bash docker run --rm -v "$PWD":/work/gluten:z apache/gluten:centos-9-jdk8 bash -c ' source /opt/rh/gcc-toolset-12/enable cd /work/gluten ./dev/builddeps-veloxbe.sh --run_setup_script=OFF --build_arrow=OFF ./build/mvn clean install -Pbackends-velox -Pspark-3.5 -DskipTests ' ``` 3. Fetch Spark 3.5.5: ```bash curl -sSL https://archive.apache.org/dist/spark/spark-3.5.5/spark-3.5.5-bin-hadoop3.tgz | tar -xz ``` 4. Write the input table and the query: ```bash cat > write_input.sql <<'EOF' CREATE TABLE t USING parquet AS SELECT 1 AS id, array(named_struct('a', 1), named_struct('a', CAST(NULL AS INT))) AS arr, CAST(NULL AS INT) AS c UNION ALL SELECT 2, array(named_struct('a', 2), CAST(NULL AS STRUCT<a: INT>)), 3; EOF cat > query.sql <<'EOF' SELECT id, filter(arr, x -> x IS NOT NULL), filter(arr, x -> x.a IS NOT NULL), filter(arr, x -> c IS NOT NULL) FROM t ORDER BY id; EXPLAIN SELECT filter(arr, x -> x IS NOT NULL), filter(arr, x -> x.a IS NOT NULL), filter(arr, x -> c IS NOT NULL) FROM t; EOF ``` 5. Write the table with plain Spark, then query it with Gluten: ```bash docker run --rm -v "$PWD":/work/gluten:z apache/gluten:centos-9-jdk8 bash -c ' cd /work/gluten rm -rf metastore_db spark-warehouse spark-3.5.5-bin-hadoop3/bin/spark-sql --master "local[2]" -f write_input.sql 2>/dev/null BUNDLE=$(ls package/target/gluten-velox-bundle-spark3.5_2.12-linux_amd64-*.jar) spark-3.5.5-bin-hadoop3/bin/spark-sql --master "local[2]" \ --jars "$BUNDLE" --driver-class-path "$BUNDLE" \ --conf spark.plugins=org.apache.gluten.GlutenPlugin \ --conf spark.memory.offHeap.enabled=true \ --conf spark.memory.offHeap.size=4g \ --conf spark.shuffle.manager=org.apache.spark.shuffle.sort.ColumnarShuffleManager \ -f query.sql 2>/dev/null ' ``` On the base commit, all three filters keep every non-null element, in the native plan: ``` 1 [{"a":1},{"a":null}] [{"a":1},{"a":null}] [{"a":1},{"a":null}] 2 [{"a":2}] [{"a":2}] [{"a":2}] == Physical Plan == VeloxColumnarToRow +- ^(1) ProjectExecTransformer [filter(arr#1, lambdafunction(isnotnull(lambda x#19), lambda x#19, false)) AS filter(arr, lambdafunction((namedlambdavariable() IS NOT NULL), namedlambdavariable()))#22, filter(arr#1, lambdafunction(isnotnull(lambda x#20.a), lambda x#20, false)) AS filter(arr, lambdafunction((namedlambdavariable().a IS NOT NULL), namedlambdavariable()))#23, filter(arr#1, lambdafunction(isnotnull(c#2), lambda x#21, false)) AS filter(arr, lambdafunction((c IS NOT NULL), namedlambdavariable()))#24] +- ^(1) FileFileSourceScanExecTransformer parquet spark_catalog.default.t[arr#1,c#2] Batched: true, DataFilters: [], Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/work/gluten/spark-warehouse/t], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<arr:array<struct<a:int>>,c:int> NativeFilters: [] ``` On this pull request's branch, rerunning steps 2 and 5 returns Spark's rows, still in the native plan: ``` 1 [{"a":1},{"a":null}] [{"a":1}] [] 2 [{"a":2}] [{"a":2}] [{"a":2},null] == Physical Plan == VeloxColumnarToRow +- ^(1) ProjectExecTransformer [filter(arr#1, lambdafunction(isnotnull(lambda x#19), lambda x#19, false)) AS filter(arr, lambdafunction((namedlambdavariable() IS NOT NULL), namedlambdavariable()))#22, filter(arr#1, lambdafunction(isnotnull(lambda x#20.a), lambda x#20, false)) AS filter(arr, lambdafunction((namedlambdavariable().a IS NOT NULL), namedlambdavariable()))#23, filter(arr#1, lambdafunction(isnotnull(c#2), lambda x#21, false)) AS filter(arr, lambdafunction((c IS NOT NULL), namedlambdavariable()))#24] +- ^(1) FileFileSourceScanExecTransformer parquet spark_catalog.default.t[arr#1,c#2] Batched: true, DataFilters: [], Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/work/gluten/spark-warehouse/t], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<arr:array<struct<a:int>>,c:int> NativeFilters: [] ``` Spark without Gluten, which is step 5's second command without the `--jars`, `--driver-class-path` and `--conf` options, returns the same rows: ``` 1 [{"a":1},{"a":null}] [{"a":1}] [] 2 [{"a":2}] [{"a":2}] [{"a":2},null] == Physical Plan == Project [filter(arr#1, lambdafunction(isnotnull(lambda x#19), lambda x#19, false)) AS filter(arr, lambdafunction((namedlambdavariable() IS NOT NULL), namedlambdavariable()))#22, filter(arr#1, lambdafunction(isnotnull(lambda x#20.a), lambda x#20, false)) AS filter(arr, lambdafunction((namedlambdavariable().a IS NOT NULL), namedlambdavariable()))#23, filter(arr#1, lambdafunction(isnotnull(c#2), lambda x#21, false)) AS filter(arr, lambdafunction((c IS NOT NULL), namedlambdavariable()))#24] +- *(1) ColumnarToRow +- FileScan parquet spark_catalog.default.t[arr#1,c#2] Batched: true, DataFilters: [], Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/work/gluten/spark-warehouse/t], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<arr:array<struct<a:int>>,c:int> ``` ### Automated tests `filter with is not null on a struct field or an outer column`, added to [ScalarFunctionsValidateSuite](https://github.com/apache/gluten/blob/cdf5292884895126825e4349938f2da3959fc812/backends-velox/src/test/scala/org/apache/gluten/functions/ScalarFunctionsValidateSuite.scala#L1058), compares Gluten with Spark on a struct-field predicate and an outer-column predicate. With the base commit's code, the new test fails: ```bash ./build/mvn -o -B test -Pbackends-velox -Pspark-3.5 -pl backends-velox -Dspotless.check.skip=true -Dscalastyle.skip=true -Dcheckstyle.skip=true -Dtest=none -DfailIfNoTests=false -Dsuites="org.apache.gluten.functions.ScalarFunctionsValidateSuite filter" ``` ``` - map_filter - filter - filter with is not null on a struct field or an outer column *** FAILED *** … !== Correct Answer - 2 == == Gluten Answer - 2 == struct<> struct<> ![ArrayBuffer([1])] [ArrayBuffer([1], [null])] [ArrayBuffer([2])] [ArrayBuffer([2])] (GlutenQueryTest.scala:460) … Tests: succeeded 2, failed 1, canceled 0, ignored 0, pending 0 ``` With this change: ```bash ./build/mvn -o -B test -Pbackends-velox -Pspark-3.5 -pl backends-velox -Dspotless.check.skip=true -Dscalastyle.skip=true -Dcheckstyle.skip=true -Dtest=none -DfailIfNoTests=false -DwildcardSuites=org.apache.gluten.functions.ScalarFunctionsValidateSuite ``` ``` ScalarFunctionsValidateSuite: … - array_compact … - filter - filter with is not null on a struct field or an outer column … Tests: succeeded 91, failed 0, canceled 0, ignored 0, pending 0 All tests passed. ``` The Bolt backend change compiles: ```bash ./build/mvn -o -B compile -Pbackends-bolt -Pspark-3.5 -pl backends-bolt -Dspotless.check.skip=true -Dscalastyle.skip=true ``` ``` [INFO] BUILD SUCCESS ``` ## Was this patch authored or co-authored using generative AI tooling? Generated-by: Co-authored with Claude Code (claude-opus-5-5) -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
