dwsmith1983 commented on code in PR #6785: URL: https://github.com/apache/datafusion-comet/pull/6785#discussion_r4393052532
########## 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: > Preserve the window sort for checkpoint probes; forced inner join followed by dense_rank still fails with missing PARTITION BY ordering. Reproduced, on 3.5 and 4.1. The strategy plans the hash join before Spark places sorts, and the join reports the checkpoint's `(k, v)` order, so Spark drops the window's sort. Comet can't verify a checkpoint's order, so the join declares none and the native window refused its input. On main the sort survived, since the forced join replaced a sort-merge join whose order covers only `k`. c20eb20f8 restores that sort on the native side: the window now sorts its input by the partition keys and then the order keys whenever its input doesn't show every partition key as ordered. That is the same condition the native window rejects, so plans that work today are unchanged. Restoring Spark's sort instead would mean changing the order the hash join reports to Spark, which the outer join order tests depend on. `a native window above an inner hash join over a checkpointed streamed side runs` covers the forced, `shuffle_hash` and `broadcast` joins with `dense_rank` and `sum(v) OVER (PARTITION BY k)`, and `a native window over a sorted checkpoint runs without Spark's sort` covers a window directly over the checkpoint; both failed before. `window_input_order.rs` covers when the sort is added on the native side. This complements #7256, which declares the window's input order when Comet can trust it. With both in, a trusted input gets the declaration and no sort, and only an order Comet can't check is sorted. -- 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]
