rdtr commented on code in PR #12151:
URL: https://github.com/apache/gluten/pull/12151#discussion_r3452406870
##########
backends-velox/src/main/scala/org/apache/gluten/extension/BloomFilterMightContainJointRewriteRule.scala:
##########
@@ -21,63 +21,38 @@ import
org.apache.gluten.expression.VeloxBloomFilterMightContain
import org.apache.gluten.expression.aggregate.VeloxBloomFilterAggregate
import org.apache.spark.sql.SparkSession
-import org.apache.spark.sql.catalyst.expressions.{BinaryExpression,
BloomFilterMightContain, Expression}
-import
org.apache.spark.sql.catalyst.expressions.aggregate.{BloomFilterAggregate,
TypedImperativeAggregate}
+import org.apache.spark.sql.catalyst.expressions.BloomFilterMightContain
+import
org.apache.spark.sql.catalyst.expressions.aggregate.{AggregateExpression,
BloomFilterAggregate}
+import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
import org.apache.spark.sql.catalyst.rules.Rule
-import org.apache.spark.sql.execution.SparkPlan
-case class BloomFilterMightContainJointRewriteRule(
- spark: SparkSession,
- isBloomFilterStatFunction: Boolean)
- extends Rule[SparkPlan] {
- override def apply(plan: SparkPlan): SparkPlan = {
- if (isBloomFilterStatFunction ||
!GlutenConfig.get.enableNativeBloomFilter) {
+/**
+ * Optimizer rule that rewrites `BloomFilterAggregate` ->
`VeloxBloomFilterAggregate` and
+ * `BloomFilterMightContain` -> `VeloxBloomFilterMightContain` at the logical
plan level.
+ *
+ * Running as an optimizer rule ensures the substitution is captured in the
`originalPlan` snapshot
+ * that
[[org.apache.gluten.extension.columnar.heuristic.ExpandFallbackPolicy]] uses
when promoting
+ * an individual stage fallback to a whole-stage AQE fallback. This guarantees
that both sides of
+ * the bloom-filter pair always produce and consume the same byte format,
regardless of whether
+ * stages fall back to JVM execution after AQE re-planning.
+ */
+case class BloomFilterMightContainJointRewriteRule(spark: SparkSession)
+ extends Rule[LogicalPlan] {
+
+ override def apply(plan: LogicalPlan): LogicalPlan = {
+ if (!GlutenConfig.get.enableNativeBloomFilter) {
Review Comment:
It looks like the early-exit `if (!GlutenConfig.get.enableNativeBloomFilter)
return plan` is untested. Can we add a test case?
##########
gluten-ut/test/src/test/scala/org/apache/gluten/sql/GlutenBloomFilterFallbackSuite.scala:
##########
@@ -0,0 +1,151 @@
+/*
+ * 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.
+ */
+package org.apache.gluten.sql
+
+import org.apache.gluten.backendsapi.BackendsApiManager
+import org.apache.gluten.config.GlutenConfig
+import org.apache.gluten.execution.WholeStageTransformerSuite
+
+import org.apache.spark.sql.catalyst.FunctionIdentifier
+import org.apache.spark.sql.catalyst.expressions.BloomFilterMightContain
+import org.apache.spark.sql.catalyst.expressions.ExpressionInfo
+import org.apache.spark.sql.catalyst.expressions.aggregate.BloomFilterAggregate
+import org.apache.spark.sql.internal.SQLConf
+
+/**
+ * Regression tests for https://github.com/apache/gluten/issues/12013.
+ *
+ * Verifies that `BloomFilterMightContainJointRewriteRule`, registered as a
`Rule[LogicalPlan]` via
+ * `injectOptimizerRule`, correctly handles whole-stage AQE fallback scenarios
where one or both
+ * bloom-filter stages revert to vanilla Spark execution.
+ */
+class GlutenBloomFilterFallbackSuite extends WholeStageTransformerSuite {
+ protected val resourcePath: String = null
+ protected val fileFormat: String = null
+
+ import testImplicits._
+
+ private val funcIdBloomFilterAgg = FunctionIdentifier("bloom_filter_agg")
+ private val funcIdMightContain = FunctionIdentifier("might_contain")
+
+ override def beforeAll(): Unit = {
+ super.beforeAll()
+ spark.sessionState.functionRegistry.registerFunction(
+ funcIdBloomFilterAgg,
+ new ExpressionInfo(classOf[BloomFilterAggregate].getName,
"bloom_filter_agg"),
+ args =>
+ args.size match {
+ case 1 => new BloomFilterAggregate(args(0))
+ case 2 => new BloomFilterAggregate(args(0), args(1))
+ case 3 => new BloomFilterAggregate(args(0), args(1), args(2))
+ case _ => throw new IllegalArgumentException("bloom_filter_agg
requires 1-3 arguments")
+ }
+ )
+ spark.sessionState.functionRegistry.registerFunction(
+ funcIdMightContain,
+ new ExpressionInfo(classOf[BloomFilterMightContain].getName,
"might_contain"),
+ args => BloomFilterMightContain(args(0), args(1)))
+ }
+
+ override def afterAll(): Unit = {
+ spark.sessionState.functionRegistry.dropFunction(funcIdBloomFilterAgg)
+ spark.sessionState.functionRegistry.dropFunction(funcIdMightContain)
+ super.afterAll()
+ }
+
+ private val veloxBloomFilterMaxNumBits = 4194304L
+
+ // GLUTEN-12013: only filter stage falls back (threshold=2).
+ // bloom_filter_agg subquery runs natively and produces Velox-format bytes;
+ // the filter stage falls back via ExpandFallbackPolicy. The optimizer-level
+ // substitution ensures the fallback plan still uses
VeloxBloomFilterMightContain
+ // so the JVM filter can read Velox-format bytes without a version mismatch.
+ testWithMinSparkVersion(
+ "GLUTEN-12013: bloom_filter_agg whole-stage fallback does not corrupt
bloom filter bytes",
+ "3.3") {
+ if
(BackendsApiManager.getSettings.requireBloomFilterAggMightContainJointFallback())
{
+ val table = "bloom_filter_test"
+ val numEstimatedItems = 5000000L
+ val sqlString =
+ s"""
+ |SELECT col positive_membership_test
+ |FROM $table
+ |WHERE might_contain(
+ | (SELECT bloom_filter_agg(col,
+ | cast($numEstimatedItems as long),
+ | cast($veloxBloomFilterMaxNumBits as long))
+ | FROM $table), col)
+ |""".stripMargin
+ withTempView(table) {
+ (Seq(Long.MinValue, 0, Long.MaxValue) ++ (1L to 200000L))
+ .toDF("col")
+ .createOrReplaceTempView(table)
+ // Threshold=2: FilterExec fallback cost=2 triggers whole-stage
fallback; agg cost=1
+ // does not, so Stage 0 runs natively. ANSI off keeps agg cost at 1
on Spark 4.0+.
+ withSQLConf(
+ GlutenConfig.COLUMNAR_FILTER_ENABLED.key -> "false",
+ GlutenConfig.COLUMNAR_WHOLESTAGE_FALLBACK_THRESHOLD.key -> "2",
+ SQLConf.ANSI_ENABLED.key -> "false"
+ ) {
+ val df = spark.sql(sqlString)
+ // Must not throw: java.io.IOException: Unexpected Bloom filter
version number.
+ assert(df.collect().length == 200003)
+ }
+ }
+ }
+ }
+
+ // GLUTEN-12013: both stages fall back (threshold=1).
+ // Stage 0's inherent transition cost of 1 meets the threshold so
ExpandFallbackPolicy
+ // promotes it to a whole-stage fallback too. Stage 0 produces Spark-format
bytes;
Review Comment:
Is this comment accurate? The optimizer rule unconditionally rewrites both
sides to Velox variants (when `enableNativeBloomFilter=true`). Falling back to
row-mode doesn't revert the expressions. Both sides still use Velox variants
and produce/consume Velox-format bytes.
Please let me know if I am missing anything. Thanks!
##########
gluten-ut/test/src/test/scala/org/apache/gluten/sql/GlutenBloomFilterFallbackSuite.scala:
##########
@@ -0,0 +1,151 @@
+/*
+ * 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.
+ */
+package org.apache.gluten.sql
+
+import org.apache.gluten.backendsapi.BackendsApiManager
+import org.apache.gluten.config.GlutenConfig
+import org.apache.gluten.execution.WholeStageTransformerSuite
+
+import org.apache.spark.sql.catalyst.FunctionIdentifier
+import org.apache.spark.sql.catalyst.expressions.BloomFilterMightContain
+import org.apache.spark.sql.catalyst.expressions.ExpressionInfo
+import org.apache.spark.sql.catalyst.expressions.aggregate.BloomFilterAggregate
+import org.apache.spark.sql.internal.SQLConf
+
+/**
+ * Regression tests for https://github.com/apache/gluten/issues/12013.
+ *
+ * Verifies that `BloomFilterMightContainJointRewriteRule`, registered as a
`Rule[LogicalPlan]` via
+ * `injectOptimizerRule`, correctly handles whole-stage AQE fallback scenarios
where one or both
+ * bloom-filter stages revert to vanilla Spark execution.
+ */
+class GlutenBloomFilterFallbackSuite extends WholeStageTransformerSuite {
+ protected val resourcePath: String = null
+ protected val fileFormat: String = null
+
+ import testImplicits._
+
+ private val funcIdBloomFilterAgg = FunctionIdentifier("bloom_filter_agg")
+ private val funcIdMightContain = FunctionIdentifier("might_contain")
+
+ override def beforeAll(): Unit = {
+ super.beforeAll()
+ spark.sessionState.functionRegistry.registerFunction(
+ funcIdBloomFilterAgg,
+ new ExpressionInfo(classOf[BloomFilterAggregate].getName,
"bloom_filter_agg"),
+ args =>
+ args.size match {
+ case 1 => new BloomFilterAggregate(args(0))
+ case 2 => new BloomFilterAggregate(args(0), args(1))
+ case 3 => new BloomFilterAggregate(args(0), args(1), args(2))
+ case _ => throw new IllegalArgumentException("bloom_filter_agg
requires 1-3 arguments")
+ }
+ )
+ spark.sessionState.functionRegistry.registerFunction(
+ funcIdMightContain,
+ new ExpressionInfo(classOf[BloomFilterMightContain].getName,
"might_contain"),
+ args => BloomFilterMightContain(args(0), args(1)))
+ }
+
+ override def afterAll(): Unit = {
+ spark.sessionState.functionRegistry.dropFunction(funcIdBloomFilterAgg)
+ spark.sessionState.functionRegistry.dropFunction(funcIdMightContain)
+ super.afterAll()
+ }
+
+ private val veloxBloomFilterMaxNumBits = 4194304L
+
+ // GLUTEN-12013: only filter stage falls back (threshold=2).
+ // bloom_filter_agg subquery runs natively and produces Velox-format bytes;
+ // the filter stage falls back via ExpandFallbackPolicy. The optimizer-level
+ // substitution ensures the fallback plan still uses
VeloxBloomFilterMightContain
+ // so the JVM filter can read Velox-format bytes without a version mismatch.
+ testWithMinSparkVersion(
+ "GLUTEN-12013: bloom_filter_agg whole-stage fallback does not corrupt
bloom filter bytes",
+ "3.3") {
+ if
(BackendsApiManager.getSettings.requireBloomFilterAggMightContainJointFallback())
{
+ val table = "bloom_filter_test"
+ val numEstimatedItems = 5000000L
+ val sqlString =
+ s"""
+ |SELECT col positive_membership_test
+ |FROM $table
+ |WHERE might_contain(
+ | (SELECT bloom_filter_agg(col,
+ | cast($numEstimatedItems as long),
+ | cast($veloxBloomFilterMaxNumBits as long))
+ | FROM $table), col)
+ |""".stripMargin
+ withTempView(table) {
+ (Seq(Long.MinValue, 0, Long.MaxValue) ++ (1L to 200000L))
+ .toDF("col")
+ .createOrReplaceTempView(table)
+ // Threshold=2: FilterExec fallback cost=2 triggers whole-stage
fallback; agg cost=1
+ // does not, so Stage 0 runs natively. ANSI off keeps agg cost at 1
on Spark 4.0+.
+ withSQLConf(
+ GlutenConfig.COLUMNAR_FILTER_ENABLED.key -> "false",
+ GlutenConfig.COLUMNAR_WHOLESTAGE_FALLBACK_THRESHOLD.key -> "2",
+ SQLConf.ANSI_ENABLED.key -> "false"
+ ) {
+ val df = spark.sql(sqlString)
+ // Must not throw: java.io.IOException: Unexpected Bloom filter
version number.
+ assert(df.collect().length == 200003)
+ }
+ }
+ }
+ }
+
+ // GLUTEN-12013: both stages fall back (threshold=1).
+ // Stage 0's inherent transition cost of 1 meets the threshold so
ExpandFallbackPolicy
+ // promotes it to a whole-stage fallback too. Stage 0 produces Spark-format
bytes;
+ // Stage 1 must use vanilla BloomFilterMightContain (not
VeloxBloomFilterMightContain)
+ // to read them. The optimizer rule rewrites both sides together so when
Stage 0 falls
+ // back the whole query uses the vanilla variants end-to-end.
+ testWithMinSparkVersion(
+ "GLUTEN-12013: bloom_filter_agg whole-stage fallback when both stages fall
back",
+ "3.3") {
+ if
(BackendsApiManager.getSettings.requireBloomFilterAggMightContainJointFallback())
{
+ val table = "bloom_filter_test"
+ val numEstimatedItems = 5000000L
+ val sqlString =
+ s"""
+ |SELECT col positive_membership_test
+ |FROM $table
+ |WHERE might_contain(
+ | (SELECT bloom_filter_agg(col,
+ | cast($numEstimatedItems as long),
+ | cast($veloxBloomFilterMaxNumBits as long))
+ | FROM $table), col)
+ |""".stripMargin
+ withTempView(table) {
+ (Seq(Long.MinValue, 0, Long.MaxValue) ++ (1L to 200000L))
+ .toDF("col")
+ .createOrReplaceTempView(table)
+ // Threshold=1: both stages fall back. Stage 0 produces Spark-format
bytes.
+ withSQLConf(
+ GlutenConfig.COLUMNAR_FILTER_ENABLED.key -> "false",
+ GlutenConfig.COLUMNAR_WHOLESTAGE_FALLBACK_THRESHOLD.key -> "1",
+ SQLConf.ANSI_ENABLED.key -> "false"
+ ) {
+ val df = spark.sql(sqlString)
+ // Must not throw: java.io.IOException: Unexpected Bloom filter
version number.
+ assert(df.collect().length == 200003)
Review Comment:
This PR rewrites `BloomFilterAggregate` and `BloomFilterMightContain` to
their Velox JVM variants unconditionally at the logical level. When stages fall
back to row-mode, the Velox variants still execute via JNI -- they aren't
reverted to vanilla are they?
The assertion proves the query doesn't throw, but doesn't verify the fix is
actually active. If both sides accidentally reverted to vanilla variants, the
query would also succeed (consistent Spark-format bytes) without the native
bloom filter path being exercised. Could we also assert that
`VeloxBloomFilterMightContain` / `VeloxBloomFilterAggregate` are present in the
executed plan? For threshold=2, the filter stage should contain
`VeloxBloomFilterMightContain`. For threshold=1, both Velox variants should be
present.
--
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]
