sunchao commented on code in PR #7117: URL: https://github.com/apache/datafusion-comet/pull/7117#discussion_r4328511918
########## spark/src/main/scala/org/apache/comet/rules/MakeDateEvaluation.scala: ########## @@ -0,0 +1,201 @@ +/* + * 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.CometConf +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") + + private def enabled(plan: SparkPlan): Boolean = + CometConf.COMET_EXEC_MAKE_DATE_OPERATOR_FALLBACK_ENABLED.get(plan.conf) + + def reason(plan: SparkPlan): Option[String] = + if (enabled(plan)) plan.getTagValue(sparkEvaluation) else None + + private def containsMakeDate(expressions: Seq[Expression]): Boolean = + expressions.exists(_.exists(_.isInstanceOf[MakeDate])) + + private def copyLogicalLinks(from: SparkPlan, to: SparkPlan): Unit = { + if (from ne to) { + // setLogicalLink promotes inherited links to direct links and rewrites descendants. + // Keep each tag's role and retain the fallback plan's links when the source has none. + Seq(SparkPlan.LOGICAL_PLAN_TAG, SparkPlan.LOGICAL_PLAN_INHERITED_TAG).foreach { tag => + from.getTagValue(tag).foreach(to.setTagValue(tag, _)) + } + } + } + + 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) + copyLogicalLinks(plan, restored) + 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 = { + if (!enabled(plan)) return plan + + 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 _: MapPartitionsExec => Some("a mapPartitions function") + 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") Review Comment: [P2] [P2] Preserve deferred dates across conditional projections Could `childProjectionDemand` also account for lazy `ProjectExec` consumers? With a Parquet table `t(y INT)` containing `6000000` and `2024`, ANSI disabled, whole-stage codegen enabled, and default optimizer rules, run: ```sql SELECT r, IF(r < 0, dt, DATE '2024-01-01') AS d FROM ( SELECT rand(7) AS r, make_date(y, 1, 1) AS dt FROM t ) ``` Spark returns `2024-01-01` for both rows without evaluating `dt`. The reused nondeterministic `r` keeps the projections separate. This head leaves both projections native and raises `ARITHMETIC_OVERFLOW` while evaluating the child. The expression guard cannot detect this because the parent contains only the `dt` attribute. Before this PR, the non-ANSI native constructor returned NULL for that overflow, so the query succeeded. Please preserve the Spark projection pipeline when a parent conditionally consumes a child’s date output, and add this cross-projection regression. The equivalent `COALESCE` shape fails too. Evidence: Reproduced on Spark 4.1.3/JDK 21 after recompiling all changed Scala sources at the reviewed SHA. Both IF and COALESCE variants failed with SparkArithmeticException/ARITHMETIC_OVERFLOW under AQE=false and AQE=true, while Spark returned two selected-date rows. Setting spark.comet.exec.project.enabled=false retained CometNativeScan and made all four controls match Spark. Spark ProjectExec.usedInputs defers single-use attributes, and doConsume forces nondeterministic outputs separately. The base kernel used i32::try_from(days).ok(), establishing the introduced non-ANSI failure. Disposable reproducer and results: /tmp/7117-10255-verified-gp3d9ynn/MakeDateLazyProjectVerification.scala and /tmp/7117-10255-verified-gp3d9ynn/probe.log. -- 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]
