dwsmith1983 commented on code in PR #6785: URL: https://github.com/apache/datafusion-comet/pull/6785#discussion_r4371605376
########## spark/src/main/scala/org/apache/comet/rules/CometShuffledHashJoinStrategy.scala: ########## @@ -0,0 +1,175 @@ +/* + * 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.comet.rules + +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.RowOrdering +import org.apache.spark.sql.catalyst.optimizer.{BuildLeft, BuildRight, BuildSide} +import org.apache.spark.sql.catalyst.planning.ExtractEquiJoinKeys +import org.apache.spark.sql.catalyst.plans.{ExistenceJoin, LeftSemi} +import org.apache.spark.sql.catalyst.plans.logical.{Join, JoinHint, JoinStrategyHint, LogicalPlan} +import org.apache.spark.sql.execution.{SparkPlan, SparkStrategy} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, LogicalQueryStage} +import org.apache.spark.sql.execution.joins.{ShuffledHashJoinExec, SortMergeJoinExec} +import org.apache.spark.sql.internal.SQLConf + +import org.apache.comet.CometConf +import org.apache.comet.CometSparkSessionExtensions.{isCometLoaded, withFallbackReason, withInfo} +import org.apache.comet.shims.ShimJoinSelection + +/** + * Plans a ShuffledHashJoinExec in place of a SortMergeJoinExec, before EnsureRequirements and + * RemoveRedundantSorts, so the sorts above the join survive. Adapted from Apache Gluten. + */ +case class CometShuffledHashJoinStrategy(session: SparkSession) + extends SparkStrategy + with ShimJoinSelection { + + override def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match { + case join: Join if isEnabled(join) => planJoin(join) + case _ => Nil + } + + private def sqlConf: SQLConf = session.sessionState.conf + + // Planner strategies run with Comet off and before CometRule, which applies plan-only mode. + private def isEnabled(join: Join): Boolean = + CometConf.COMET_FORCE_SHJ.get(sqlConf) && + CometConf.COMET_EXEC_ENABLED.get(sqlConf) && + !CometConf.COMET_EXPLAIN_PLAN_ONLY_ENABLED.get(sqlConf) && + !join.isStreaming && + isCometLoaded(sqlConf) + + private def planJoin(join: Join): Seq[SparkPlan] = join match { + case ExtractEquiJoinKeys(joinType, leftKeys, rightKeys, condition, _, left, right, hint) + if !isLeftToSpark(join, hint) && canSortMergeJoin(joinType) && + RowOrdering.isOrderable(leftKeys) && hashJoinSupportedShim(leftKeys, rightKeys) => + buildSide(join) match { + case Some(BuildRight) if joinType == LeftSemi || joinType.isInstanceOf[ExistenceJoin] => + // LeftSemi https://github.com/apache/datafusion-comet/issues/2667 + // ExistenceJoin https://github.com/apache/datafusion-comet/issues/2697 + planBySpark(join) { smj => + withFallbackReason(smj, declineReason(s"BuildRight with $joinType is not supported")) + } + case Some(side) => + buildSideOverLimit(join, side) match { + case Some(reason) => + // Not a fallback: the kept sort-merge join still runs natively. + planBySpark(join)(smj => withInfo(smj, declineReason(reason))) + case None => + Seq( + ShuffledHashJoinExec( + leftKeys, + rightKeys, + joinType, + side, + condition, + planLater(left), Review Comment: `CometHashJoin` now sends the streamed side's order for every join that reports one, not only the two outer shapes, whenever `mayReportFalseOrder` finds nothing under the streamed side that may have moved its rows. An inner join keeps its probe order either way, so declaring it moves no rows, and the native window above now sees the order. When the order cannot be checked or its first key cannot be expressed, an inner join declares nothing and stays native, as before. `an inner hash join declares a sorted cache's order for a native window above it` runs your shape with forcing on and with `shuffle_hash` and `broadcast` hints. `an inner hash join declares the part of the streamed order Comet can express` and `an inner hash join over a checkpointed streamed side declares no order` cover the two cases that declare nothing, and `inner_hash_join_declares_streamed_order_on_probe_input` in `planner.rs` covers the native side for both build sides. A window directly over the sorted cache, with no join in between, fails the same way on main. That needs the window itself to carry the order, so I filed it separately: #7223. -- 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]
