yadavay-amzn commented on code in PR #56101: URL: https://github.com/apache/spark/pull/56101#discussion_r3660217728
########## sql/core/src/main/scala/org/apache/spark/sql/execution/joins/BroadcastNearestByJoinExec.scala: ########## @@ -0,0 +1,172 @@ +/* + * 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.joins + +import java.util.{Comparator, PriorityQueue => JPriorityQueue} + +import org.apache.spark.SparkException +import org.apache.spark.rdd.RDD +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions._ +import org.apache.spark.sql.catalyst.plans.{InnerLike, JoinType, LeftOuter, NearestByDirection, NearestByDistance} +import org.apache.spark.sql.catalyst.plans.physical._ +import org.apache.spark.sql.catalyst.util.TypeUtils +import org.apache.spark.sql.execution.{ExplainUtils, SparkPlan} +import org.apache.spark.sql.execution.metric.SQLMetrics + +/** + * Heap entry storing an index into the broadcast array alongside its ranking value. + * Using a case class with primitive `Int` field avoids boxing that `(Int, Any)` tuples incur. + */ +private[joins] case class HeapEntry(index: Int, rankingValue: Any) + +/** + * Physical operator for NearestByJoin that avoids materializing the full cross product. + * For each left row, iterates all broadcast right rows maintaining a bounded priority + * queue of size k, then emits the top-k matches directly. + * + * The right side is fully broadcast to all partitions. This operator only fires when + * the right side fits within [[SQLConf.AUTO_BROADCASTJOIN_THRESHOLD]]. For right tables + * exceeding this threshold, the existing cross-product + aggregate rewrite is used as + * fallback. Tie-breaking among equal ranking values is non-deterministic (matches the + * existing rewrite behavior). + */ +case class BroadcastNearestByJoinExec( + left: SparkPlan, + right: SparkPlan, + joinType: JoinType, + numResults: Int, + rankingExpression: Expression, + direction: NearestByDirection) extends BaseJoinExec { + + override def condition: Option[Expression] = None + override def leftKeys: Seq[Expression] = Seq.empty + override def rightKeys: Seq[Expression] = Seq.empty + + override def simpleStringWithNodeId(): String = { + val opId = ExplainUtils.getOpId(this) + s"$nodeName $joinType k=$numResults $direction ($opId)".trim + } + + override def output: Seq[Attribute] = joinType match { + case _: InnerLike | LeftOuter => + left.output.map(_.withNullability(true)) ++ right.output.map(_.withNullability(true)) + case other => + throw SparkException.internalError( + s"$nodeName does not support join type: $other") + } + + override lazy val metrics = Map( + "numOutputRows" -> SQLMetrics.createMetric(sparkContext, "number of output rows"), + "streamedRows" -> SQLMetrics.createMetric(sparkContext, "number of left rows processed")) + + override def requiredChildDistribution: Seq[Distribution] = + UnspecifiedDistribution :: BroadcastDistribution(IdentityBroadcastMode) :: Nil + + override def outputPartitioning: Partitioning = left.outputPartitioning + + override def outputOrdering: Seq[SortOrder] = Nil + + protected override def doExecute(): RDD[InternalRow] = { + val broadcastedRight = right.executeBroadcast[Array[InternalRow]]() + val numOutput = longMetric("numOutputRows") + val streamedRowsMetric = longMetric("streamedRows") + val localJoinType = joinType + val k = numResults + val isDistance = direction == NearestByDistance + val leftOutput = left.output + val rightOutput = right.output + val rankExpr = rankingExpression + val allOutput = output + val ordering = TypeUtils.getInterpretedOrdering(rankExpr.dataType) + + left.execute().mapPartitionsInternal { leftIter => + val rightRows = broadcastedRight.value + if (rightRows.isEmpty && localJoinType != LeftOuter) { + Iterator.empty + } else { + val joinedRow = new JoinedRow + val rankingProj = UnsafeProjection.create( Review Comment: You're right, thanks. Switched to `mapPartitionsWithIndexInternal` and call `rankingProj.initialize(partitionIndex)` before eval, so a `rand()` ranking no longer throws. Added a test for it. ########## sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/RewriteNearestByJoin.scala: ########## @@ -72,7 +73,14 @@ object RewriteNearestByJoin extends Rule[LogicalPlan] { private lazy val random = new scala.util.Random() def apply(plan: LogicalPlan): LogicalPlan = plan.transformUp { - case j @ NearestByJoin(left, right, joinType, _, numResults, rankingExpression, direction) => + case j @ NearestByJoin(left, right, joinType, _, numResults, rankingExpression, direction) + // The optimizer checks both the config flag AND the broadcast threshold to decide + // whether to skip the rewrite. This mixes optimizer/planner concerns but is necessary: + // if we only checked the config flag, a NearestByJoin node with right side exceeding + // the broadcast threshold would reach the planner unrewritten, and no strategy would + // handle it (NearestByJoinSelection returns Nil for large right sides), causing a + // planning failure. The alternative (two-pass approach) is deferred to future work. + if !NearestByJoin.canBroadcastRight(j, SQLConf.get) => Review Comment: The gate is needed in the rewrite: without it a large-right `NearestByJoin` reaches `NearestByJoinSelection`, which returns `Nil` (a planning failure). And since estimates only shrink through optimization, reading the early size is conservative: it can miss a broadcast but never picks one wrongly, and the planner re-checks with refined stats. Kept it as-is, but happy to revisit if the missed-opportunity case matters. ########## sql/core/src/main/scala/org/apache/spark/sql/execution/joins/BroadcastNearestByJoinExec.scala: ########## @@ -0,0 +1,149 @@ +/* + * 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.joins + +import java.util.{PriorityQueue => JPriorityQueue} + +import org.apache.spark.SparkException +import org.apache.spark.rdd.RDD +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions._ +import org.apache.spark.sql.catalyst.plans.{InnerLike, JoinType, LeftOuter, NearestByDirection, NearestByDistance} +import org.apache.spark.sql.catalyst.plans.physical._ +import org.apache.spark.sql.execution.{BinaryExecNode, SparkPlan} +import org.apache.spark.sql.execution.metric.SQLMetrics +import org.apache.spark.sql.types.DoubleType + +/** + * Physical operator for NearestByJoin that avoids materializing the full cross product. + * For each left row, iterates all broadcast right rows maintaining a bounded priority + * queue of size k, then emits the top-k matches directly. + * + * The right side is fully broadcast to all partitions. This operator only fires when + * the right side fits within [[SQLConf.AUTO_BROADCAST_JOIN_THRESHOLD]]. For right tables Review Comment: Done, used the fully-qualified `[[org.apache.spark.sql.internal.SQLConf.autoBroadcastJoinThreshold]]` so it resolves without an import. -- 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]
