Skip to content

Commit

Permalink
[GLUTEN-8168] Add pre-projections for join condition (#8185)
Browse files Browse the repository at this point in the history
* refactor

* add new rule

* fixed
  • Loading branch information
lgbo-ustc authored Dec 12, 2024
1 parent e12db45 commit 3eb30a5
Show file tree
Hide file tree
Showing 8 changed files with 293 additions and 84 deletions.
Original file line number Diff line number Diff line change
Expand Up @@ -350,6 +350,13 @@ object CHBackendSettings extends BackendSettingsApi with Logging {
)
}

def enablePreProjectionForJoinConditions(): Boolean = {
SparkEnv.get.conf.getBoolean(
CHConf.runtimeConfig("enable_pre_projection_for_join_conditions"),
defaultValue = true
)
}

// If the partition keys are high cardinality, the aggregation method is slower.
def enableConvertWindowGroupLimitToAggregate(): Boolean = {
SparkEnv.get.conf.getBoolean(
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -111,6 +111,7 @@ object CHRuleApi {
c.session)))
injector.injectPostTransform(c => InsertTransitions.create(c.outputsColumnar, CHBatch))
injector.injectPostTransform(c => RemoveDuplicatedColumns.apply(c.session))
injector.injectPostTransform(c => AddPreProjectionForHashJoin.apply(c.session))

// Gluten columnar: Fallback policies.
injector.injectFallbackPolicy(
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,202 @@
/*
* 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.extension

import org.apache.gluten.backendsapi.clickhouse.CHBackendSettings
import org.apache.gluten.execution._
import org.apache.gluten.utils._

import org.apache.spark.internal.Logging
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.catalyst.expressions._
import org.apache.spark.sql.catalyst.expressions.NamedExpression
import org.apache.spark.sql.catalyst.rules.Rule
import org.apache.spark.sql.execution._

import scala.collection.mutable

/*
* It will add a projection before hash join for
* 1. The join keys which are not attribute references.
* 2. If the join could rewrite into multiple join contiditions, also replace all the join keys with
* attribute references.
*
* JoinUtils.createPreProjectionIfNeeded has add a pre-projection for join keys, but it doesn't
* handle the post filter.
*/
case class AddPreProjectionForHashJoin(session: SparkSession)
extends Rule[SparkPlan]
with PullOutProjectHelper
with Logging {
override def apply(plan: SparkPlan): SparkPlan = {
if (!CHBackendSettings.enablePreProjectionForJoinConditions) {
return plan
}

plan.transformUp {
case hashJoin: CHShuffledHashJoinExecTransformer =>
val leftReplacedExpressions = new mutable.HashMap[Expression, NamedExpression]
val rightReplacedExpressions = new mutable.HashMap[Expression, NamedExpression]

val newLeftKeys = hashJoin.leftKeys.map {
case e => replaceExpressionWithAttribute(e, leftReplacedExpressions, false, false)
}

val newRightKeys = hashJoin.rightKeys.map {
case e => replaceExpressionWithAttribute(e, rightReplacedExpressions, false, false)
}

val newCondition = replaceExpressionInCondition(
hashJoin.condition,
hashJoin.left,
leftReplacedExpressions,
hashJoin.right,
rightReplacedExpressions)
val leftProjectExprs =
eliminateProjectList(hashJoin.left.outputSet, leftReplacedExpressions.values.toSeq)
val rightProjectExprs =
eliminateProjectList(hashJoin.right.outputSet, rightReplacedExpressions.values.toSeq)
val newHashJoin = hashJoin.copy(
leftKeys = newLeftKeys,
rightKeys = newRightKeys,
condition = newCondition,
left = if (leftReplacedExpressions.size > 0) {
ProjectExecTransformer(leftProjectExprs, hashJoin.left)
} else { hashJoin.left },
right = if (rightReplacedExpressions.size > 0) {
ProjectExecTransformer(rightProjectExprs, hashJoin.right)
} else { hashJoin.right }
)
if (leftReplacedExpressions.size > 0 || rightReplacedExpressions.size > 0) {
ProjectExecTransformer(hashJoin.output, newHashJoin)
} else {
newHashJoin
}
}
}

private def replaceExpressionInCondition(
condition: Option[Expression],
leftPlan: SparkPlan,
leftReplacedExpressions: mutable.HashMap[Expression, NamedExpression],
rightPlan: SparkPlan,
rightReplacedExpressions: mutable.HashMap[Expression, NamedExpression])
: Option[Expression] = {
if (!condition.isDefined) {
return condition
}

def replaceExpression(
e: Expression,
exprMap: mutable.HashMap[Expression, NamedExpression]): Expression = {
e match {
case or @ Or(left, right) =>
Or(replaceExpression(left, exprMap), replaceExpression(right, exprMap))
case and @ And(left, right) =>
And(replaceExpression(left, exprMap), replaceExpression(right, exprMap))
case equalTo @ EqualTo(left, right) =>
EqualTo(replaceExpression(left, exprMap), replaceExpression(right, exprMap))
case _ =>
exprMap.getOrElseUpdate(e.canonicalized, null) match {
case null => e
case ne: NamedExpression => ne.toAttribute
}
}
}

val leftExprs = new mutable.ArrayBuffer[Expression]
val rightExps = new mutable.ArrayBuffer[Expression]
if (isMultipleOrEqualsCondition(condition.get, leftPlan, leftExprs, rightPlan, rightExps)) {
leftExprs.foreach {
e => replaceExpressionWithAttribute(e, leftReplacedExpressions, false, false)
}
rightExps.foreach {
e => replaceExpressionWithAttribute(e, rightReplacedExpressions, false, false)
}
Some(replaceExpression(condition.get, leftReplacedExpressions ++ rightReplacedExpressions))
} else {
condition
}
}

private def isMultipleOrEqualsCondition(
e: Expression,
leftPlan: SparkPlan,
leftExpressions: mutable.ArrayBuffer[Expression],
rightPlan: SparkPlan,
rightExpressions: mutable.ArrayBuffer[Expression]): Boolean = {
def splitIntoOrExpressions(e: Expression, result: mutable.ArrayBuffer[Expression]): Unit = {
e match {
case Or(left, right) =>
splitIntoOrExpressions(left, result)
splitIntoOrExpressions(right, result)
case _ =>
result += e
}
}
def splitIntoAndExpressions(e: Expression, result: mutable.ArrayBuffer[Expression]): Unit = {
e match {
case And(left, right) =>
splitIntoAndExpressions(left, result)
splitIntoAndExpressions(right, result)
case _ =>
result += e
}
}

val leftOutputSet = leftPlan.outputSet
val rightOutputSet = rightPlan.outputSet
val orExpressions = new mutable.ArrayBuffer[Expression]()
splitIntoOrExpressions(e, orExpressions)
orExpressions.foreach {
orExpression =>
val andExpressions = new mutable.ArrayBuffer[Expression]()
splitIntoAndExpressions(orExpression, andExpressions)
andExpressions.foreach {
e =>
if (!e.isInstanceOf[EqualTo]) {
return false;
}
val equalExpr = e.asInstanceOf[EqualTo]
val leftPos = if (equalExpr.left.references.subsetOf(leftOutputSet)) {
leftExpressions += equalExpr.left
0
} else if (equalExpr.left.references.subsetOf(rightOutputSet)) {
rightExpressions += equalExpr.left
1
} else {
return false
}
val rightPos = if (equalExpr.right.references.subsetOf(leftOutputSet)) {
leftExpressions += equalExpr.right
0
} else if (equalExpr.right.references.subsetOf(rightOutputSet)) {
rightExpressions += equalExpr.right
1
} else {
return false
}

// they should come from different side
if (leftPos == rightPos) {
return false
}
}
}
true
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -105,4 +105,40 @@ class GlutenClickHouseJoinSuite extends GlutenClickHouseWholeStageTransformerSui
}
}

test("GLUTEN-8168 eliminate non-attribute expressions in join keys and condition") {
sql("create table tj1 (a int, b int, c int, d int) using parquet")
sql("create table tj2 (a int, b int, c int, d int) using parquet")
sql("insert into tj1 values (1, 2, 3, 4), (2, 2, 4, 5), (3, 4, 5, 4), (4, 5, 3, 7)")
sql("insert into tj2 values (1, 2, 3, 4), (2, 2, 4, 5), (3, 4, 5, 4), (4, 5, 3, 7)")
compareResultsAgainstVanillaSpark(
"""
|SELECT t1.*, t2.*
|FROM tj1 t1 LEFT JOIN tj2 t2
|ON t1.a = t2.a AND (t1.b = t2.b or t1.c + 1 = t2.c) ORDER BY t1.a, t2.a
|""".stripMargin,
true,
{ _ => }
)
compareResultsAgainstVanillaSpark(
"""
|SELECT t1.*, t2.*
|FROM tj1 t1 LEFT JOIN tj2 t2
|ON t1.a = t2.a AND (t1.b = t2.b or t1.c = t2.c) ORDER BY t1.a, t2.a
|""".stripMargin,
true,
{ _ => }
)
compareResultsAgainstVanillaSpark(
"""
|SELECT t1.*, t2.*
|FROM tj1 t1 LEFT JOIN tj2 t2
|ON t1.a = t2.a + 1 AND (t1.b + 1 = t2.b or t1.c = t2.c + 1) ORDER BY t1.a, t2.a
|""".stripMargin,
true,
{ _ => }
)
sql("drop table if exists tj1")
sql("drop table if exists tj2")
}

}
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,6 @@ import scala.collection.mutable
class GlutenClickHouseTPCHBucketSuite
extends GlutenClickHouseTPCHAbstractSuite
with AdaptiveSparkPlanHelper {

override protected val tablesPath: String = basePath + "/tpch-data-ch"
override protected val tpchQueries: String = rootPath + "queries/tpch-queries-ch"
override protected val queriesResults: String = rootPath + "bucket-queries-output"
Expand All @@ -48,6 +47,7 @@ class GlutenClickHouseTPCHBucketSuite
.set("spark.sql.autoBroadcastJoinThreshold", "-1") // for test bucket join
.set("spark.sql.adaptive.enabled", "true")
.set("spark.gluten.sql.columnar.backend.ch.shuffle.hash.algorithm", "sparkMurmurHash3_32")
.setCHConfig("enable_pre_projection_for_join_conditions", "false")
.setCHConfig("enable_grace_aggregate_spill_test", "true")
}

Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -51,6 +51,7 @@ class GlutenClickHouseTPCHMetricsSuite extends GlutenClickHouseTPCHAbstractSuite
.set("spark.sql.autoBroadcastJoinThreshold", "10MB")
.setCHConfig("logger.level", "error")
.setCHSettings("input_format_parquet_max_block_size", parquetMaxBlockSize)
.setCHConfig("enable_pre_projection_for_join_conditions", "false")
.setCHConfig("enable_streaming_aggregating", true)
}
// scalastyle:on line.size.limit
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ class GlutenClickHouseTPCHParquetBucketSuite
.set("spark.sql.autoBroadcastJoinThreshold", "-1") // for test bucket join
.set("spark.sql.adaptive.enabled", "true")
.set("spark.gluten.sql.columnar.backend.ch.shuffle.hash.algorithm", "sparkMurmurHash3_32")
.setCHConfig("enable_pre_projection_for_join_conditions", "false")
.setCHConfig("enable_grace_aggregate_spill_test", "true")
}

Expand Down
Loading

0 comments on commit 3eb30a5

Please sign in to comment.