cloud-fan commented on a change in pull request #26434: [SPARK-29544] [SQL] optimize skewed partition based on data size URL: https://github.com/apache/spark/pull/26434#discussion_r344829344
########## File path: sql/core/src/main/scala/org/apache/spark/sql/execution/adaptive/OptimizeSkewedPartitions.scala ########## @@ -0,0 +1,261 @@ +/* + * 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.spark.sql.execution.adaptive + +import scala.collection.mutable +import scala.concurrent.duration.Duration + +import org.apache.spark.MapOutputStatistics +import org.apache.spark.rdd.RDD +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions.Attribute +import org.apache.spark.sql.catalyst.plans._ +import org.apache.spark.sql.catalyst.plans.physical.{Partitioning, UnknownPartitioning} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution._ +import org.apache.spark.sql.execution.joins.SortMergeJoinExec +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.util.ThreadUtils + +case class OptimizeSkewedPartitions(conf: SQLConf) extends Rule[SparkPlan] { + + private val supportedJoinTypes = + Inner :: Cross :: LeftSemi :: LeftAnti :: LeftOuter :: RightOuter :: Nil + + /** + * A partition is considered as a skewed partition if its size is larger than the median + * partition size * spark.sql.adaptive.skewedPartitionFactor and also larger than + * spark.sql.adaptive.skewedPartitionSizeThreshold. + */ + private def isSkewed( + stats: MapOutputStatistics, + partitionId: Int, + medianSize: Long): Boolean = { + val size = stats.bytesByPartitionId(partitionId) + size > medianSize * conf.adaptiveSkewedFactor && + size > conf.adaptiveSkewedSizeThreshold + } + + private def medianSize(stats: MapOutputStatistics): Long = { + val bytesLen = stats.bytesByPartitionId.length + val bytes = stats.bytesByPartitionId.sorted + if (bytes(bytesLen / 2) > 0) bytes(bytesLen / 2) else 1 + } + + /** + * To equally divide n elements into m buckets, basically each bucket should have n/m elements, + * for the remaining n%m elements, add one more element to the first n%m buckets each. Returns + * a sequence with length numBuckets and each value represents the start index of each bucket. + */ + def equallyDivide(numElements: Int, numBuckets: Int): Seq[Int] = { + val elementsPerBucket = numElements / numBuckets + val remaining = numElements % numBuckets + val splitPoint = (elementsPerBucket + 1) * remaining + (0 until remaining).map(_ * (elementsPerBucket + 1)) ++ + (remaining until numBuckets).map(i => splitPoint + (i - remaining) * elementsPerBucket) + } + + /** + * We split the partition into several splits. Each split reads the data from several map outputs + * ranging from startMapId to endMapId(exclusive). This method calculates the split number and + * the startMapId for all splits. + */ + private def estimateMapIdStartIndices( + stage: QueryStageExec, + partitionId: Int, + medianSize: Long): Array[Int] = { + val metrics = getStatistics(stage) + val size = metrics.bytesByPartitionId(partitionId) + val factor = size / medianSize + val numMappers = getShuffleStage(stage). + plan.shuffleDependency.rdd.partitions.length + val numSplits = Math.min(conf.adaptiveSkewedMaxSplits, + Math.min(factor.toInt, numMappers)) + equallyDivide(numMappers, numSplits).toArray + } + + private def getShuffleStage(queryStage: QueryStageExec): ShuffleQueryStageExec = { + queryStage match { + case stage: ShuffleQueryStageExec => stage + case ReusedQueryStageExec(_, stage: ShuffleQueryStageExec, _) => stage + } + } + + private def getStatistics(queryStage: QueryStageExec): MapOutputStatistics = { + val shuffleStage = queryStage match { + case stage: ShuffleQueryStageExec => stage + case ReusedQueryStageExec(_, stage: ShuffleQueryStageExec, _) => stage + } + val metrics = shuffleStage.mapOutputStatisticsFuture + assert(metrics.isCompleted, "ShuffleQueryStageExec should already be ready") + ThreadUtils.awaitResult(metrics, Duration.Zero) + } + + /** + * Base optimization support check: the join type is supported and plan statistics is available. + * Note that for some join types(like left outer), whether a certain partition can be optimized + * also depends on the filed isSkewAndSupportsSplit. + */ + private def supportOptimization( + joinType: JoinType, + leftStage: QueryStageExec, + rightStage: QueryStageExec): Boolean = { + val joinTypeSupported = supportedJoinTypes.contains(joinType) + val shuffleStageCheck = ShuffleQueryStageExec.isShuffleQueryStageExec(leftStage) && + ShuffleQueryStageExec.isShuffleQueryStageExec(rightStage) + val statisticsReady: Boolean = if (shuffleStageCheck) { + getStatistics(leftStage) != null && getStatistics(rightStage) != null + } else false + + joinTypeSupported && statisticsReady + } + + private def supportSplitOnLeftPartition(joinType: JoinType) = joinType != RightOuter + + private def supportSplitOnRightPartition(joinType: JoinType) = { + joinType != LeftOuter && joinType != LeftSemi && joinType != LeftAnti + } + + def handleSkewJoin(plan: SparkPlan): SparkPlan = plan.transformUp { + case smj @ SortMergeJoinExec(leftKeys, rightKeys, joinType, condition, + SortExec(_, _, left: QueryStageExec, _), + SortExec(_, _, right: QueryStageExec, _)) + if supportOptimization(joinType, left, right) => + val leftStats = getStatistics(left) + val rightStats = getStatistics(right) + val numPartitions = leftStats.bytesByPartitionId.length + + val leftMedSize = medianSize(leftStats) + val rightMedSize = medianSize(rightStats) + logInfo(s"HandlingSkewedJoin left medSize: ($leftMedSize)" + + s" right medSize ($rightMedSize)") + logInfo(s"left bytes Max : ${leftStats.bytesByPartitionId.max}") + logInfo(s"right bytes Max : ${rightStats.bytesByPartitionId.max}") + + val skewedPartitions = mutable.HashSet[Int]() + val subJoins = mutable.ArrayBuffer[SparkPlan](smj) + for (partitionId <- 0 until numPartitions) { + val isLeftSkew = isSkewed(leftStats, partitionId, leftMedSize) + val isRightSkew = isSkewed(rightStats, partitionId, rightMedSize) + val isSkewAndSupportsSplit = + (isLeftSkew && supportSplitOnLeftPartition(joinType)) || + (isRightSkew && supportSplitOnRightPartition(joinType)) + + if (isSkewAndSupportsSplit) { + skewedPartitions += partitionId + val leftMapIdStartIndices = if (isLeftSkew && supportSplitOnLeftPartition(joinType)) { + estimateMapIdStartIndices(left, partitionId, leftMedSize) + } else { + Array(0) + } + val rightMapIdStartIndices = if (isRightSkew && supportSplitOnRightPartition(joinType)) { + estimateMapIdStartIndices(right, partitionId, rightMedSize) + } else { + Array(0) + } + + for (i <- 0 until leftMapIdStartIndices.length; + j <- 0 until rightMapIdStartIndices.length) { + val leftEndMapId = if (i == leftMapIdStartIndices.length - 1) { + getShuffleStage(left).plan.shuffleDependency.rdd.partitions.length + } else { + leftMapIdStartIndices(i + 1) + } + val rightEndMapId = if (j == rightMapIdStartIndices.length - 1) { + getShuffleStage(right). + plan.shuffleDependency.rdd.partitions.length + } else { + rightMapIdStartIndices(j + 1) + } + // For the skewed partition, we set the id of shuffle query stage to -1. + // And skip this shuffle query stage optimization in 'ReduceNumShufflePartitions' rule. + val leftSkewedReader = + SkewedShuffleReaderExec(getShuffleStage(left).copy(id = -1), + partitionId, leftMapIdStartIndices(i), leftEndMapId) + + val rightSkewedReader = + SkewedShuffleReaderExec(getShuffleStage(right).copy(id = -1), + partitionId, rightMapIdStartIndices(j), rightEndMapId) + + subJoins += + SortMergeJoinExec(leftKeys, rightKeys, joinType, condition, + leftSkewedReader, rightSkewedReader) + } + } + } + logInfo(s"skewed partition number is ${skewedPartitions.size}") + if (skewedPartitions.size > 0) { + getShuffleStage(left).skewedPartitions = skewedPartitions + getShuffleStage(right).skewedPartitions = skewedPartitions + UnionExec(subJoins.toList) + } else { + smj + } + } + + override def apply(plan: SparkPlan): SparkPlan = { + if (!conf.adaptiveSkewedJoinEnabled) { + return plan + } + + def collectShuffleStages(plan: SparkPlan): Seq[ShuffleQueryStageExec] = plan match { + case _: LocalShuffleReaderExec => Nil + case _: SkewedShuffleReaderExec => Nil + case stage: ShuffleQueryStageExec => Seq(stage) + case ReusedQueryStageExec(_, stage: ShuffleQueryStageExec, _) => Seq(stage) + case _ => plan.children.flatMap(collectShuffleStages) + } + + val shuffleStages = collectShuffleStages(plan) + + if (shuffleStages.length == 2) { + // Currently we only support handling skewed join for 2 table join. + handleSkewJoin(plan) + } else { + plan + } + } +} + +case class SkewedShuffleReaderExec( Review comment: This is different from local/coalesced shuffle reader as it reads only one reduce partition. Maybe better to call it `PostShufflePartitionReader` ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
