|
| 1 | +/* |
| 2 | + * Licensed to the Apache Software Foundation (ASF) under one or more |
| 3 | + * contributor license agreements. See the NOTICE file distributed with |
| 4 | + * this work for additional information regarding copyright ownership. |
| 5 | + * The ASF licenses this file to You under the Apache License, Version 2.0 |
| 6 | + * (the "License"); you may not use this file except in compliance with |
| 7 | + * the License. You may obtain a copy of the License at |
| 8 | + * |
| 9 | + * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | + * |
| 11 | + * Unless required by applicable law or agreed to in writing, software |
| 12 | + * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | + * See the License for the specific language governing permissions and |
| 15 | + * limitations under the License. |
| 16 | + */ |
| 17 | + |
| 18 | +package org.apache.spark.sql.execution.bucketing |
| 19 | + |
| 20 | +import scala.annotation.tailrec |
| 21 | + |
| 22 | +import org.apache.spark.sql.catalyst.catalog.BucketSpec |
| 23 | +import org.apache.spark.sql.catalyst.expressions.Expression |
| 24 | +import org.apache.spark.sql.catalyst.optimizer.{BuildLeft, BuildRight} |
| 25 | +import org.apache.spark.sql.catalyst.plans.physical.{HashPartitioning, Partitioning} |
| 26 | +import org.apache.spark.sql.catalyst.rules.Rule |
| 27 | +import org.apache.spark.sql.execution.{FileSourceScanExec, FilterExec, ProjectExec, SparkPlan} |
| 28 | +import org.apache.spark.sql.execution.joins.{BaseJoinExec, ShuffledHashJoinExec, SortMergeJoinExec} |
| 29 | +import org.apache.spark.sql.internal.SQLConf |
| 30 | + |
| 31 | +/** |
| 32 | + * This rule coalesces one side of the `SortMergeJoin` and `ShuffledHashJoin` |
| 33 | + * if the following conditions are met: |
| 34 | + * - Two bucketed tables are joined. |
| 35 | + * - Join keys match with output partition expressions on their respective sides. |
| 36 | + * - The larger bucket number is divisible by the smaller bucket number. |
| 37 | + * - COALESCE_BUCKETS_IN_JOIN_ENABLED is set to true. |
| 38 | + * - The ratio of the number of buckets is less than the value set in |
| 39 | + * COALESCE_BUCKETS_IN_JOIN_MAX_BUCKET_RATIO. |
| 40 | + */ |
| 41 | +case class CoalesceBucketsInJoin(conf: SQLConf) extends Rule[SparkPlan] { |
| 42 | + private def updateNumCoalescedBucketsInScan( |
| 43 | + plan: SparkPlan, |
| 44 | + numCoalescedBuckets: Int): SparkPlan = { |
| 45 | + plan transformUp { |
| 46 | + case f: FileSourceScanExec => |
| 47 | + f.copy(optionalNumCoalescedBuckets = Some(numCoalescedBuckets)) |
| 48 | + } |
| 49 | + } |
| 50 | + |
| 51 | + private def updateNumCoalescedBuckets( |
| 52 | + join: BaseJoinExec, |
| 53 | + numLeftBuckets: Int, |
| 54 | + numRightBucket: Int, |
| 55 | + numCoalescedBuckets: Int): BaseJoinExec = { |
| 56 | + if (numCoalescedBuckets != numLeftBuckets) { |
| 57 | + val leftCoalescedChild = |
| 58 | + updateNumCoalescedBucketsInScan(join.left, numCoalescedBuckets) |
| 59 | + join match { |
| 60 | + case j: SortMergeJoinExec => j.copy(left = leftCoalescedChild) |
| 61 | + case j: ShuffledHashJoinExec => j.copy(left = leftCoalescedChild) |
| 62 | + } |
| 63 | + } else { |
| 64 | + val rightCoalescedChild = |
| 65 | + updateNumCoalescedBucketsInScan(join.right, numCoalescedBuckets) |
| 66 | + join match { |
| 67 | + case j: SortMergeJoinExec => j.copy(right = rightCoalescedChild) |
| 68 | + case j: ShuffledHashJoinExec => j.copy(right = rightCoalescedChild) |
| 69 | + } |
| 70 | + } |
| 71 | + } |
| 72 | + |
| 73 | + private def isCoalesceSHJStreamSide( |
| 74 | + join: ShuffledHashJoinExec, |
| 75 | + numLeftBuckets: Int, |
| 76 | + numRightBucket: Int, |
| 77 | + numCoalescedBuckets: Int): Boolean = { |
| 78 | + if (numCoalescedBuckets == numLeftBuckets) { |
| 79 | + join.buildSide != BuildRight |
| 80 | + } else { |
| 81 | + join.buildSide != BuildLeft |
| 82 | + } |
| 83 | + } |
| 84 | + |
| 85 | + def apply(plan: SparkPlan): SparkPlan = { |
| 86 | + if (!conf.coalesceBucketsInJoinEnabled) { |
| 87 | + return plan |
| 88 | + } |
| 89 | + |
| 90 | + plan transform { |
| 91 | + case ExtractJoinWithBuckets(join, numLeftBuckets, numRightBuckets) |
| 92 | + if math.max(numLeftBuckets, numRightBuckets) / math.min(numLeftBuckets, numRightBuckets) <= |
| 93 | + conf.coalesceBucketsInJoinMaxBucketRatio => |
| 94 | + val numCoalescedBuckets = math.min(numLeftBuckets, numRightBuckets) |
| 95 | + join match { |
| 96 | + case j: SortMergeJoinExec => |
| 97 | + updateNumCoalescedBuckets(j, numLeftBuckets, numRightBuckets, numCoalescedBuckets) |
| 98 | + case j: ShuffledHashJoinExec |
| 99 | + // Only coalesce the buckets for shuffled hash join stream side, |
| 100 | + // to avoid OOM for build side. |
| 101 | + if isCoalesceSHJStreamSide(j, numLeftBuckets, numRightBuckets, numCoalescedBuckets) => |
| 102 | + updateNumCoalescedBuckets(j, numLeftBuckets, numRightBuckets, numCoalescedBuckets) |
| 103 | + case other => other |
| 104 | + } |
| 105 | + case other => other |
| 106 | + } |
| 107 | + } |
| 108 | +} |
| 109 | + |
| 110 | +/** |
| 111 | + * An extractor that extracts `SortMergeJoinExec` and `ShuffledHashJoin`, |
| 112 | + * where both sides of the join have the bucketed tables, |
| 113 | + * are consisted of only the scan operation, |
| 114 | + * and numbers of buckets are not equal but divisible. |
| 115 | + */ |
| 116 | +object ExtractJoinWithBuckets { |
| 117 | + @tailrec |
| 118 | + private def hasScanOperation(plan: SparkPlan): Boolean = plan match { |
| 119 | + case f: FilterExec => hasScanOperation(f.child) |
| 120 | + case p: ProjectExec => hasScanOperation(p.child) |
| 121 | + case _: FileSourceScanExec => true |
| 122 | + case _ => false |
| 123 | + } |
| 124 | + |
| 125 | + private def getBucketSpec(plan: SparkPlan): Option[BucketSpec] = { |
| 126 | + plan.collectFirst { |
| 127 | + case f: FileSourceScanExec if f.relation.bucketSpec.nonEmpty && |
| 128 | + f.optionalNumCoalescedBuckets.isEmpty => |
| 129 | + f.relation.bucketSpec.get |
| 130 | + } |
| 131 | + } |
| 132 | + |
| 133 | + /** |
| 134 | + * The join keys should match with expressions for output partitioning. Note that |
| 135 | + * the ordering does not matter because it will be handled in `EnsureRequirements`. |
| 136 | + */ |
| 137 | + private def satisfiesOutputPartitioning( |
| 138 | + keys: Seq[Expression], |
| 139 | + partitioning: Partitioning): Boolean = { |
| 140 | + partitioning match { |
| 141 | + case HashPartitioning(exprs, _) if exprs.length == keys.length => |
| 142 | + exprs.forall(e => keys.exists(_.semanticEquals(e))) |
| 143 | + case _ => false |
| 144 | + } |
| 145 | + } |
| 146 | + |
| 147 | + private def isApplicable(j: BaseJoinExec): Boolean = { |
| 148 | + (j.isInstanceOf[SortMergeJoinExec] || |
| 149 | + j.isInstanceOf[ShuffledHashJoinExec]) && |
| 150 | + hasScanOperation(j.left) && |
| 151 | + hasScanOperation(j.right) && |
| 152 | + satisfiesOutputPartitioning(j.leftKeys, j.left.outputPartitioning) && |
| 153 | + satisfiesOutputPartitioning(j.rightKeys, j.right.outputPartitioning) |
| 154 | + } |
| 155 | + |
| 156 | + private def isDivisible(numBuckets1: Int, numBuckets2: Int): Boolean = { |
| 157 | + val (small, large) = (math.min(numBuckets1, numBuckets2), math.max(numBuckets1, numBuckets2)) |
| 158 | + // A bucket can be coalesced only if the bigger number of buckets is divisible by the smaller |
| 159 | + // number of buckets because bucket id is calculated by modding the total number of buckets. |
| 160 | + numBuckets1 != numBuckets2 && large % small == 0 |
| 161 | + } |
| 162 | + |
| 163 | + def unapply(plan: SparkPlan): Option[(BaseJoinExec, Int, Int)] = { |
| 164 | + plan match { |
| 165 | + case j: BaseJoinExec if isApplicable(j) => |
| 166 | + val leftBucket = getBucketSpec(j.left) |
| 167 | + val rightBucket = getBucketSpec(j.right) |
| 168 | + if (leftBucket.isDefined && rightBucket.isDefined && |
| 169 | + isDivisible(leftBucket.get.numBuckets, rightBucket.get.numBuckets)) { |
| 170 | + Some(j, leftBucket.get.numBuckets, rightBucket.get.numBuckets) |
| 171 | + } else { |
| 172 | + None |
| 173 | + } |
| 174 | + case _ => None |
| 175 | + } |
| 176 | + } |
| 177 | +} |
0 commit comments