ajsquared commented on code in PR #7117: URL: https://github.com/apache/datafusion-comet/pull/7117#discussion_r4316841742
########## spark/src/main/scala/org/apache/comet/rules/MakeDateEvaluation.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.catalyst.expressions.{Expression, MakeDate, SortOrder} +import org.apache.spark.sql.catalyst.trees.TreeNodeTag +import org.apache.spark.sql.comet._ +import org.apache.spark.sql.comet.execution.shuffle.CometShuffleExchangeExec +import org.apache.spark.sql.execution._ +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, QueryStageExec} +import org.apache.spark.sql.execution.aggregate.{BaseAggregateExec, HashAggregateExec, ObjectHashAggregateExec} +import org.apache.spark.sql.execution.exchange.{Exchange, ReusedExchangeExec} +import org.apache.spark.sql.execution.joins.BaseJoinExec + +import org.apache.comet.shims.ShimCometWindowGroupLimit + +/** Keeps make_date on Spark's row path where native batching can change whether it throws. */ +private[rules] object MakeDateEvaluation { + private val sparkEvaluation = TreeNodeTag[String]("comet.makeDateSparkEvaluation") + + def reason(plan: SparkPlan): Option[String] = plan.getTagValue(sparkEvaluation) + + private def containsMakeDate(expressions: Seq[Expression]): Boolean = + expressions.exists(_.exists(_.isInstanceOf[MakeDate])) + + private def nativeBroadcastInput(plan: SparkPlan): Option[SparkPlan] = plan match { + case _: CometBroadcastExchangeExec => Some(plan) + case stage: BroadcastQueryStageExec => nativeBroadcastInput(stage.plan).map(_ => stage) + case reuse: ReusedExchangeExec => nativeBroadcastInput(reuse.child).map(_ => reuse) + case sink: CometSinkPlaceHolder => nativeBroadcastInput(sink.child) + case wrapper: CometScanWrapper => nativeBroadcastInput(wrapper.originalPlan) + case _ => None + } + + private def restoreBroadcast(plan: SparkPlan, reason: String): SparkPlan = + nativeBroadcastInput(plan) + .map { input => + // Spark joins need a HashedRelation (or rows for a nested-loop join), not Arrow buffers. + // The existing transition decodes the cached broadcast and applies its Spark broadcast + // mode, including through AQE/reuse wrappers, without re-executing the build input. + val restored = CometColumnarToRowExec(input) + restored.setTagValue(sparkEvaluation, reason) + plan.logicalLink.foreach(restored.setLogicalLink) + restored + } + .getOrElse(plan) + + /** + * Run before bottom-up conversion and join rewriting. Reasons propagate back to the consuming + * operator, keeping the row pipeline together instead of inserting an eager Arrow batch between + * a throwing projection and its lazy consumer. Exchanges materialize their input independently; + * local sorts and aggregates do so only after their consumer requests input. + */ + def preserve(plan: SparkPlan): SparkPlan = { + def visit( + node: SparkPlan, + rowDemand: Option[String], + projectionDemand: Option[String]): (SparkPlan, Option[String]) = { + node match { + // Materialized stages are independent producers. Their plans have already been guarded. + case _: QueryStageExec | _: ReusedExchangeExec => return (node, None) + // These wrappers do not have the child shape of their original Spark operator. + case wrapper: CometScanWrapper => + val (child, childReason) = visit(wrapper.originalPlan, rowDemand, projectionDemand) + return if (childReason.isDefined) (child, childReason) + else (wrapper.copy(originalPlan = child), None) + case sink: CometSinkPlaceHolder => + val (child, childReason) = visit(sink.child, rowDemand, projectionDemand) + return if (childReason.isDefined) (child, childReason) + else (sink.withNewChildren(Seq(child)), None) + case _ => + } + + val original = node match { + case scan: CometScanExec => + scan.wrapped.copy( + partitionFilters = scan.partitionFilters, + dataFilters = scan.dataFilters) + case scan: CometNativeScanExec => scan.sparkFallback(Nil) + case native: CometExec => native.originalPlan + case shuffle: CometShuffleExchangeExec => shuffle.originalPlan + case broadcast: CometBroadcastExchangeExec => broadcast.originalPlan + case _ => node + } + val exchange = original.isInstanceOf[Exchange] + val join = original.isInstanceOf[BaseJoinExec] + val localMaterialization = original match { + case _: SortExec | _: HashAggregateExec | _: ObjectHashAggregateExec => true + case _: TakeOrderedAndProjectExec => true + case _ => false + } + val startsLimit = original match { + case _: CollectLimitExec | _: LocalLimitExec | _: GlobalLimitExec => true + case topK: TakeOrderedAndProjectExec => + SortOrder.orderingSatisfies(topK.child.outputOrdering, topK.sortOrder) + case _ + if ShimCometWindowGroupLimit.windowGroupLimitClass.exists(_.isInstance(original)) => + true + case _ => false + } + val childDemand = original match { + case _ if exchange => None + // A join can leave a local sort or aggregate completely unpolled. + case _ if join || rowDemand.contains("JOIN") => Some("JOIN") + case _ if startsLimit => Some("LIMIT") + case _: SampleExec => Some("SAMPLE") + case _ if localMaterialization => None + case _ => rowDemand + } + val childProjectionDemand = original match { + // Generated filters can defer deterministic projection outputs until a row is accepted. + case _: FilterExec => Some("FILTER") + case _ if exchange || localMaterialization || join => None + case _ => projectionDemand + } + val visited = node.children.map(visit(_, childDemand, childProjectionDemand)) + val childReason = visited.flatMap(_._2).headOption + val ownReason = reason(node).orElse { + if (containsMakeDate(original.expressions)) { + rowDemand + .map(boundary => s"make_date requires Spark evaluation below $boundary") + .orElse(original match { + case project: ProjectExec + if containsMakeDate(project.projectList.filter(_.deterministic)) => + projectionDemand.map(boundary => + s"make_date requires Spark evaluation below $boundary") + // Join keys/conditions and aggregate updates can be skipped inside the operator. + case _: BaseJoinExec | _: BaseAggregateExec | _: TakeOrderedAndProjectExec => + Some(s"make_date requires Spark evaluation in ${original.nodeName}") + case _ => None + }) + } else None + } + val fallbackReason = ownReason.orElse(if (exchange) None else childReason) + val children = visited.map { case (child, _) => + if (join && fallbackReason.isDefined) restoreBroadcast(child, fallbackReason.get) + else child + } + val rebuilt = node match { + case _: RowToColumnarExec | _: CometSparkToColumnarExec if childReason.isDefined => + children.head + case _: ColumnarToRowExec | _: CometColumnarToRowExec | _: CometNativeColumnarToRowExec + if childReason.isDefined && !children.head.supportsColumnar => + children.head + case native: CometExec if fallbackReason.isDefined || children != node.children => + native.sparkFallback(children) + case _ if (original ne node) && (fallbackReason.isDefined || children != node.children) => + original.withNewChildren(children) + case _ => node.withNewChildren(children) + } + node.logicalLink.foreach(rebuilt.setLogicalLink) + fallbackReason.foreach(rebuilt.setTagValue(sparkEvaluation, _)) + (rebuilt, fallbackReason) + } + + visit(plan, None, None)._1 Review Comment: [P2] Preserve row demand for RDD and mapPartitions consumers. Starting with no demand, and not recognizing MapPartitionsExec, leaves make_date projections native beneath these partial readers. With one Parquet batch containing y=2024 followed by y=6000000, df.selectExpr("make_date(y, 1, 1)").rdd.take(1) succeeds in Spark but now raises ARITHMETIC_OVERFLOW in Comet, including with ANSI disabled, because the native projection evaluates both rows. A mapPartitions(_.take(1)) consumer has the same gap. Seed demand for object-producing roots and recognize MapPartitionsExec, as tagPartiallyReadTypedDatasetOutputs already does, and add execution regressions. -- 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]
