imback82 commented on a change in pull request #29655: URL: https://github.com/apache/spark/pull/29655#discussion_r487594228
########## File path: sql/core/src/main/scala/org/apache/spark/sql/execution/exchange/OptimizeSortMergeJoinWithPartialHashDistribution.scala ########## @@ -0,0 +1,115 @@ +/* + * 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.spark.sql.execution.exchange + +import scala.collection.mutable + +import org.apache.spark.sql.catalyst.expressions.Expression +import org.apache.spark.sql.catalyst.plans.physical.{HashPartitioning, Partitioning} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{SortExec, SparkPlan} +import org.apache.spark.sql.execution.joins.SortMergeJoinExec +import org.apache.spark.sql.internal.SQLConf + +/** + * This rule removes shuffle for the sort merge join if the following conditions are met: + * - The child of ShuffleExchangeExec has HashPartitioning with the same number of partitions + * as the other side of join. + * - The child of ShuffleExchangeExec has output partitioning which has the subset of + * join keys on the respective join side. + * + * If the above conditions are met, shuffle can be eliminated for the sort merge join + * because rows are sorted before join logic is applied. + */ +case class OptimizeSortMergeJoinWithPartialHashDistribution(conf: SQLConf) extends Rule[SparkPlan] { + def apply(plan: SparkPlan): SparkPlan = { + if (!conf.optimizeSortMergeJoinWithPartialHashDistribution) { + return plan + } + + plan.transformUp { + case s @ SortMergeJoinExec(_, _, _, _, + lSort @ SortExec(_, _, + ExtractShuffleExchangeExecChild( + lChild, + lChildOutputPartitioning: HashPartitioning), + _), + rSort @ SortExec(_, _, + ExtractShuffleExchangeExecChild( + rChild, + rChildOutputPartitioning: HashPartitioning), + _), + false) if isPartialHashDistribution( + s.leftKeys, lChildOutputPartitioning, s.rightKeys, rChildOutputPartitioning) => + // Remove ShuffleExchangeExec. + s.copy(left = lSort.copy(child = lChild), right = rSort.copy(child = rChild)) + case other => other + } + } + + /* + * Returns true if both HashPartitioning have the same number of partitions and + * their partitioning expressions are a subset of their respective join keys. + */ + private def isPartialHashDistribution( + leftKeys: Seq[Expression], + leftPartitioning: HashPartitioning, + rightKeys: Seq[Expression], + rightPartitioning: HashPartitioning): Boolean = { + val mapping = leftKeyToRightKeyMapping(leftKeys, rightKeys) + (leftPartitioning.numPartitions == rightPartitioning.numPartitions) && + leftPartitioning.expressions.zip(rightPartitioning.expressions) + .forall { + case (le, re) => mapping.get(le.canonicalized) + .map(_.exists(_.semanticEquals(re))) + .getOrElse(false) + } Review comment: > If the number of buckets for `t1` is less than number of shuffle partitions, shouldn't it shuffle both sides ? (in [`EnsureRequirements`](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/exchange/EnsureRequirements.scala#L96)). So the rule kicks in here and removes both shuffles, but we shouldn't remove any shuffle here. You are right. Thanks for the catch! > I think it's unsafe if we do not shuffle both sides. `HashPartitioning(Seq(a, b))` and `HashPartitioning(Seq(b, a))` are not same thing, e.g. for tuple (a: 1, b: 2) it will be assigned to different buckets given current `Murmur3Hash` implementation. Yes, I understand they produce different hash values. However, it has the join condition `t1.a = t2.b AND t1.b = t2.a`. This rule will not be applied if the condition was `t1.a = t2.a AND t1.b = t2.b`. Please let me know if I missed something. Thanks! ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
