pingzh commented on code in PR #7283: URL: https://github.com/apache/datafusion-comet/pull/7283#discussion_r4415669692
########## spark/src/main/scala/org/apache/comet/rules/MakeDateEvaluation.scala: ########## @@ -0,0 +1,384 @@ +/* + * 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.{Add, Alias, And, Attribute, AttributeSet, BinaryComparison, Cast, ConditionalExpression, CreateArray, CreateNamedStruct, DateAdd, DateDiff, DateFormatClass, DateSub, EqualNullSafe, Expression, GetDateField, Greatest, HashExpression, IsNotNull, IsNull, Least, MakeDate, Multiply, Not, Or, SortOrder, Subtract, UnixDate} +import org.apache.spark.sql.catalyst.expressions.aggregate.{AggregateExpression, Count} +import org.apache.spark.sql.catalyst.optimizer.BuildLeft +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, BroadcastHashJoinExec, BroadcastNestedLoopJoinExec} + +import org.apache.comet.CometConf +import org.apache.comet.serde.CometMakeDate +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 containsDeferredAggregateDate(expression: Expression): Boolean = expression match { + // COUNT checks its sole nullable argument on every input row. Filters, other nullable + // arguments, and a non-nullable argument can let Spark skip evaluating the date. + case AggregateExpression(Count(Seq(child)), _, _, None, _) if child.nullable => false + case _: MakeDate => true + case _ => expression.children.exists(containsDeferredAggregateDate) + } + + private case class ProjectionDemand( + boundary: String, + deferred: AttributeSet, + // Deferred inputs whose producers need guarding even when they contain no make_date. + guardedInputs: AttributeSet, + // Date argument lineage retained until a streamed join can skip its input rows. + dateInputs: AttributeSet) + + // Attributes may stand for deferred expressions in a lower projection, not materialized values. + private def harmlessValue(expression: Expression): Boolean = + expression.references.isEmpty && CometMakeDate.canEvaluateEagerly(expression) + + /** References that a consumer requires without exposing a potentially skipped sibling. */ + private def eagerReferences(expression: Expression): AttributeSet = expression match { + case attribute: Attribute => AttributeSet(Seq(attribute)) + case _ => + val children = expression match { + case date: MakeDate => + val firstNullable = date.children.indexWhere(_.nullable) + date.children.take(if (firstNullable < 0) date.children.size else firstNullable + 1) + case conditional: ConditionalExpression => + val required = conditional.alwaysEvaluatedInputs + if (conditional.children.filterNot(required.contains).forall(harmlessValue)) required + else Nil + case _: Alias | _: Cast | _: IsNull | _: IsNotNull | _: Not | _: UnixDate | + _: GetDateField => + expression.children + case _: EqualNullSafe | _: CreateArray | _: CreateNamedStruct | _: Greatest | _: Least | + _: HashExpression[_] => + expression.children + case _: And | _: Or => + if (harmlessValue(expression.children(1))) expression.children.take(1) else Nil + case _: Add | _: Subtract | _: Multiply | _: BinaryComparison | _: DateAdd | _: DateSub | + _: DateDiff | _: DateFormatClass => + val left = expression.children.head + val right = expression.children(1) + if (!left.nullable) expression.children + // A right-hand attribute can inline a throwing expression from an earlier projection. + // Only self-contained harmless values are safe without tracing that attribute's source. + else if (harmlessValue(right)) Seq(left) + else Nil + // Unknown expressions may skip inputs or catch errors raised by their children. + case _ => Nil + } + AttributeSet(children.flatMap(eagerReferences)) + } + + 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[ProjectionDemand]): (SparkPlan, Option[String], AttributeSet) = { + node match { + // Materialized stages are independent producers. Their plans have already been guarded. + case _: QueryStageExec | _: ReusedExchangeExec => return (node, None, AttributeSet.empty) + // These wrappers do not have the child shape of their original Spark operator. + case wrapper: CometScanWrapper => + val (child, childReason, dateOutputs) = + visit(wrapper.originalPlan, rowDemand, projectionDemand) + return if (childReason.isDefined) (child, childReason, dateOutputs) + else (wrapper.copy(originalPlan = child), None, dateOutputs) + case sink: CometSinkPlaceHolder => + val (child, childReason, dateOutputs) = visit(sink.child, rowDemand, projectionDemand) + return if (childReason.isDefined) (child, childReason, dateOutputs) + else (sink.withNewChildren(Seq(child)), None, dateOutputs) + 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 evaluatesDate = containsMakeDate(original.expressions) + val exchange = original.isInstanceOf[Exchange] + val join = original.isInstanceOf[BaseJoinExec] + // Broadcast joins can inline streamed expressions into their consumer. Their build side + // materializes rows; shuffled hash and sort-merge joins also isolate the streamed input. + val streamedChildIndex = original match { + case hash: BroadcastHashJoinExec if hash.supportCodegen => + Some(if (hash.buildSide == BuildLeft) 1 else 0) + case nestedLoop: BroadcastNestedLoopJoinExec if nestedLoop.supportCodegen => + Some(if (nestedLoop.buildSide == BuildLeft) 1 else 0) + case _ => None + } + 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 filter: FilterExec => + Some( + ProjectionDemand( + "FILTER", + filter.child.outputSet, + projectionDemand.map(_.guardedInputs).getOrElse(AttributeSet.empty), + projectionDemand.map(_.dateInputs).getOrElse(AttributeSet.empty))) + case project: ProjectExec => + val deferred = projectionDemand.map(_.deferred).getOrElse(AttributeSet.empty) + val eager = project.projectList + .filter(output => !output.deterministic || !deferred.contains(output.toAttribute)) + .flatMap(eagerReferences) + // ProjectExec forces repeated inputs and nondeterministic outputs. Other inputs can + // remain inlined in a consumer's conditional, even across intervening aliases. + val skipped = project.child.outputSet -- project.usedInputs -- AttributeSet(eager) + val inheritedInputs = + projectionDemand.map(_.guardedInputs).getOrElse(AttributeSet.empty) + // A consumer can require the date but still skip another output of this projection. + val guardedOutputs = inheritedInputs ++ + (if (evaluatesDate) deferred else AttributeSet.empty) + val directDateInputs = AttributeSet(project.projectList.flatMap { output => + output.collect { case date: MakeDate => date }.flatMap(_.references) + }) + val inputsToGuard = + directDateInputs ++ AttributeSet(project.projectList.flatMap { output => + if ((output.deterministic || rowDemand.isDefined) && + guardedOutputs.contains(output.toAttribute)) { + output.references.toSeq + } else { + Nil + } + }) + // Keep argument lineage through eager aliases too. A join below this projection can + // discard an input row before this projection ever requires the argument's value. + val dateInputs = + directDateInputs ++ AttributeSet(project.projectList.flatMap { output => + if (projectionDemand.exists(_.dateInputs.contains(output.toAttribute))) { + output.references.toSeq + } else { + Nil + } + }) + // A lazy row consumer can skip even the date's required arguments on unread rows. + val guardedChildInputs = + inputsToGuard.intersect(if (rowDemand.isDefined) project.child.outputSet else skipped) + if (skipped.isEmpty && guardedChildInputs.isEmpty && dateInputs.isEmpty) { + None + } else { + Some( + ProjectionDemand( + projectionDemand.map(_.boundary).getOrElse("PROJECT"), + skipped, + guardedChildInputs, + dateInputs)) + } + case _ if exchange || localMaterialization => None + case _ => projectionDemand + } + val visited = node.children.zipWithIndex.map { case (child, index) => + val demand = if (!join) { + childProjectionDemand + } else if (streamedChildIndex.contains(index)) { + childProjectionDemand.map { demand => + // Keys and conditions force their inputs on each visited row. An outer row consumer + // can still skip that row altogether, so retain its guarded inputs in that case. + val required = if (rowDemand.isDefined) AttributeSet.empty else original.references + val streamedOutputs = original.children(index).outputSet + val forwarded = streamedOutputs -- required + demand.copy( + deferred = demand.deferred.intersect(forwarded), + guardedInputs = (demand.guardedInputs ++ demand.dateInputs).intersect(forwarded), + // A key is required only on this join's input rows. An earlier join can still + // discard rows before producing that key, so retain its argument lineage. + dateInputs = demand.dateInputs.intersect(streamedOutputs)) + } + } else { + None + } + visit(child, childDemand, demand) + } + val childReason = visited.flatMap(_._2).headOption + val childDateOutputs = AttributeSet(visited.zipWithIndex.flatMap { + case ((_, _, outputs), index) if !join || streamedChildIndex.contains(index) => outputs + case _ => AttributeSet.empty + }) + // Date provenance flows up through aliases while deferred evaluation flows down. This + // also finds a skipped sibling computed in a projection above the date's producer. + val dateDerivedOutputs = if (exchange || localMaterialization) { + AttributeSet.empty + } else { + original match { + case project: ProjectExec => + AttributeSet( + project.projectList + .filter { output => + containsMakeDate(Seq(output)) || output.references + .intersect(childDateOutputs) + .nonEmpty + } + .map(_.toAttribute)) + case _ => childDateOutputs.intersect(original.outputSet) + } + } + val ownReason = reason(node).orElse { + // An exchange evaluates its partition keys for every input before a consumer reads rows. + val rowReason = if (evaluatesDate && !exchange) { + rowDemand.map(boundary => s"make_date requires Spark evaluation below $boundary") + } else None + rowReason.orElse(original match { + case project: ProjectExec => + projectionDemand + .filter(demand => + project.projectList.exists { output => + val computation = output match { + case alias: Alias => alias.child + case other => other + } + val deferredDate = output.deterministic && + demand.deferred.contains(output.toAttribute) && containsMakeDate(Seq(output)) + val deferredComputation = (output.deterministic || rowDemand.isDefined) && + (demand.guardedInputs.contains(output.toAttribute) || + (dateDerivedOutputs.nonEmpty && demand.deferred.contains( + output.toAttribute))) && + !CometMakeDate.canEvaluateEagerly(computation) + deferredDate || deferredComputation + }) + .map(demand => s"make_date requires Spark evaluation below ${demand.boundary}") + case aggregate: BaseAggregateExec + if aggregate.expressions.exists(containsDeferredAggregateDate) => + Some(s"make_date requires Spark evaluation in ${original.nodeName}") + // Join keys/conditions and top-k outputs can be skipped inside the operator. + case _: BaseJoinExec | _: TakeOrderedAndProjectExec if evaluatesDate => + Some(s"make_date requires Spark evaluation in ${original.nodeName}") + case _ => 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) + } + copyLogicalLinks(node, rebuilt) + fallbackReason.foreach(rebuilt.setTagValue(sparkEvaluation, _)) + (rebuilt, fallbackReason, dateDerivedOutputs) + } + + // Dataset.rdd adds an object-producing root, and its caller may read only a few objects. + // EliminateSerialization can leave a typed operator, filter, or project at that root. + def producesObjects(op: SparkPlan): Boolean = op match { + case _: ObjectProducerExec => true + case _: FilterExec | _: ProjectExec => producesObjects(op.children.head) + case _ => false Review Comment: Agreed that `producesObjects` and the common boundary classification are worth sharing. I'd like to handle that in a separate follow-up to keep this PR focused on evaluation correctness. The propagation rules need to remain distinct: this walk carries JOIN demand through local sorts and aggregates because a join can leave them completely unpolled, while the typed-Dataset walk treats those operators as boundaries. The follow-up can share the classification and `producesObjects` while keeping those propagation decisions explicit and covered by both sets of tests. ########## spark/src/test/scala/org/apache/comet/rules/CometExecRuleSuite.scala: ########## @@ -151,6 +155,847 @@ class CometExecRuleSuite extends CometTestBase { }.sum } + test("make_date preserves row demand below lazy consumers, including reused native plans") { + withSQLConf( + SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false", + CometConf.COMET_EXEC_LOCAL_TABLE_SCAN_ENABLED.key -> "true") { + val input = createSparkPlan(spark, "SELECT * FROM VALUES (2024), (2025) AS t(y)") + val sample = createSparkPlan( + spark, + "SELECT * FROM (SELECT * FROM range(8)) TABLESAMPLE (50 PERCENT)").collectFirst { + case op: SampleExec => op + }.get + val consumers: Seq[(String, SparkPlan => SparkPlan)] = Seq( + ("LIMIT", child => CollectLimitExec(1, child)), + ("LIMIT", child => LocalLimitExec(1, child)), + ("LIMIT", child => GlobalLimitExec(1, child)), + ("SAMPLE", child => sample.withNewChildren(Seq(child))), + ("FILTER", child => FilterExec(GreaterThan(child.output.head, Literal(0)), child))) + for (ansi <- Seq(false, true)) { + def project: SparkPlan = ProjectExec( + Seq( + input.output.head, + Alias(MakeDate(input.output.head, Literal(1), Literal(1), ansi), "date")()), + input) + val native = applyCometExecRule(project) + assert(countOperators(native, classOf[CometProjectExec]) == 1, native.treeString) + for ((boundary, consume) <- consumers; reused <- Seq(false, true)) { + val child = if (reused) applyCometExecRule(project) else project + val converted = applyCometExecRule(consume(child)) + for (result <- Seq(converted, applyCometExecRule(converted))) { + withClue(s"ansi=$ansi reused=$reused $boundary\n${result.treeString}") { + assert(countOperators(result, classOf[ProjectExec]) == 1) + assert(countOperators(result, classOf[CometProjectExec]) == 0) + assert( + new ExtendedExplainInfo() + .getFallbackReasons(result) + .exists(_.contains(s"make_date requires Spark evaluation below $boundary"))) + } + } + } + + // These consumers evaluate their complete input independently of the outer limit. + val materializers: Seq[SparkPlan => SparkPlan] = Seq( + child => SortExec(Seq(SortOrder(child.output.head, Ascending)), false, child), + child => ShuffleExchangeExec(SinglePartition, child)) + materializers.foreach { materialize => + val result = applyCometExecRule(LocalLimitExec(1, materialize(project))) + assert(countOperators(result, classOf[CometProjectExec]) == 1, result.treeString) + assert( + !new ExtendedExplainInfo() + .getFallbackReasons(result) + .exists(_.contains("make_date requires Spark evaluation below LIMIT"))) + } + } + } + } + + test("make_date shuffle keys remain eager below a limit") { + withSQLConf( + SQLConf.ANSI_ENABLED.key -> "true", + SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false", + CometConf.COMET_SHUFFLE_ENABLED.key -> "true", + CometConf.COMET_SHUFFLE_MODE.key -> "native", + CometConf.COMET_SHUFFLE_NATIVE_HASH_PARTITIONING_ENABLED.key -> "true") { + for (invalid <- Seq(false, true)) { + val rows = Seq.fill(20)((2024, 1, 1)) :+ ((2024, if (invalid) 13 else 1, 1)) + withParquetTable(rows, "make_date_shuffle") { + val df = spark + .table("make_date_shuffle") + .repartition(8, expr("make_date(_1, _2, _3)")) + .limit(10) + val plan = stripAQEPlan(df.queryExecution.executedPlan) + val exchanges = plan.collect { case exchange: CometShuffleExchangeExec => exchange } + withClue(plan.treeString) { + assert(exchanges.size == 1) + assert(exchanges.head.shuffleType == CometNativeShuffle) + assert( + exchanges.head.originalPlan.expressions.exists(_.exists(_.isInstanceOf[MakeDate]))) + } + if (invalid) { + // LIMIT cannot hide an invalid key, even when enough valid input rows precede it. + val (sparkError, cometError) = checkSparkAnswerMaybeThrows(df) + Seq("Spark" -> sparkError, "Comet" -> cometError).foreach { case (engine, error) => + withClue(s"$engine: ") { + assert( + error.exists(e => causeChain(e).exists(_.toString.contains("MonthOfYear")))) + } + } + } else { + checkSparkAnswer(df) + } + } + } + } + } + + test("make_date join input demand reaches local sorts before hash join rewriting") { + withSQLConf( + SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false", + CometConf.COMET_EXEC_LOCAL_TABLE_SCAN_ENABLED.key -> "true", + CometConf.COMET_FORCE_SHJ.key -> "true", + CometConf.COMET_FORCE_SHJ_MAX_BUILD_SIZE.key -> "-1") { + val left = createSparkPlan(spark, "SELECT * FROM VALUES (2024), (2025) AS t(y)") + val right = createSparkPlan(spark, "SELECT * FROM VALUES (2024) AS t(y)") + val date = MakeDate(left.output.head, Literal(1), Literal(1), failOnError = false) + val projection = ProjectExec(Seq(left.output.head, Alias(date, "date")()), left) + def sort(child: SparkPlan): SparkPlan = + SortExec(Seq(SortOrder(child.output.head, Ascending)), false, child) + val join = SortMergeJoinExec( + Seq(left.output.head), + Seq(right.output.head), + Inner, + None, + sort(applyCometExecRule(projection)), + sort(right)) + val converted = applyCometExecRule(join) + for (result <- Seq(converted, applyCometExecRule(converted))) { + assert(countOperators(result, classOf[SortMergeJoinExec]) == 1, result.treeString) + assert(countOperators(result, classOf[CometProjectExec]) == 0, result.treeString) + assert(countOperators(result, classOf[ProjectExec]) == 1, result.treeString) + assert( + new ExtendedExplainInfo() + .getFallbackReasons(result) + .exists(_.contains("make_date requires Spark evaluation below JOIN"))) + } + } + } + + test("make_date remains deferred across conditional projections") { Review Comment: Agreed that the query/result and fallback-reason cases are good candidates for the SQL harness. I'd like to move those in a separate follow-up after verifying that the `rand`-based queries retain the required projection shapes under the ANSI/AQE configurations. The physical operator counts, native scan checks, broadcast reuse, and opt-out checks would stay in Scala. Both suites already run under the same Spark-profile CI matrix, so the move would improve test organization while preserving the existing version coverage. Keeping it separate also makes it easier to check that the migration retains each regression's original failure condition. -- 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]
