dbaliafroozeh commented on a change in pull request #28885: URL: https://github.com/apache/spark/pull/28885#discussion_r447072694
########## File path: sql/core/src/main/scala/org/apache/spark/sql/execution/reuse/Reuse.scala ########## @@ -0,0 +1,90 @@ +/* + * 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.reuse + +import scala.collection.mutable.Map + +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BaseSubqueryExec, ExecSubqueryExpression, ReusedSubqueryExec, SparkPlan} +import org.apache.spark.sql.execution.exchange.{Exchange, ReusedExchangeExec} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.types.StructType + +/** + * Find out duplicated exchanges and subqueries in the whole spark plan including subqueries, then + * use the same exhange or subquery for all the references. + */ +case class ReuseExchangeAndSubquery(conf: SQLConf) extends Rule[SparkPlan] { + + private class ReuseCache[T <: SparkPlan] { + // To avoid costly canonicalization of an exchange or a subquery: + // - we use its schema first to check if it can be replaced to a reused one at all + // - we insert it into the map of canonicalized plans only when at least 2 have the same schema + private val cache = Map[StructType, (T, Map[SparkPlan, T])]() + + def lookup(plan: T): T = { + val (firstSameSchemaPlan, sameResultPlans) = cache.getOrElseUpdate(plan.schema, plan -> Map()) + if (firstSameSchemaPlan.ne(plan)) { + if (sameResultPlans.isEmpty) { + sameResultPlans += firstSameSchemaPlan.canonicalized -> firstSameSchemaPlan + } + sameResultPlans.getOrElseUpdate(plan.canonicalized, plan) + } else { + plan + } + } + } + + def apply(plan: SparkPlan): SparkPlan = { + if (conf.exchangeReuseEnabled || conf.subqueryReuseEnabled) { + val exchanges = new ReuseCache[Exchange]() + val subqueries = new ReuseCache[BaseSubqueryExec]() + + def reuse(plan: SparkPlan): SparkPlan = plan.transformUp { + case exchange: Exchange if conf.exchangeReuseEnabled => + val cached = exchanges.lookup(exchange) + if (cached.ne(exchange)) { + ReusedExchangeExec(exchange.output, cached) + } else { + exchange + } + + case other => other.transformExpressionsUp { + case sub: ExecSubqueryExpression => + val subquery = reuse(sub.plan).asInstanceOf[BaseSubqueryExec] + sub.withNewPlan( + if (conf.subqueryReuseEnabled) { Review comment: Yeah, I agree `lookupOrElse` is not the best name. I was thinking more about it, so we want a method with the following semantics: Try to find an existing node with the same canonicalized form: - if no such node found, add the node to the map and return the node itself - if a node found that refers the same instance as the key, do nothing, and return the node itself - otherwise, call the given function on the retrieved node from the map and then return the result I think it's not like anything (I mean common methods) on the Map interfaces that I know of. Somehow reminds me of Java 8 `putIfAbsent` method, but with the difference it calls the passed function if it's present. On second thought, maybe we should rename the `lookup` method to `getOrElseUpdate` because it doesn't just lookup. And call the second method `applyIfPresent`. Maybe also doesn't make sense to have the second method anymore since it's very specific and hard to capture with a method name what's going on. One last thing, what was happening before if we had two exchanges in the query plan that were referring to the same instance? Were we leaving them intact or replacing it with a reuse node? Because now we just leave it as is. I'm not sure though if such a situation actually can happen to have to exchanges referring to the same instance in the query plan, if not, maybe we can remove the check and (the second method) altogether. Also maybe putting a reuse node when there are the same instances is even better, it's just a wrapper and signals the presence of the same exchange node. ---------------------------------------------------------------- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
