Github user JoshRosen commented on a diff in the pull request:
https://github.com/apache/spark/pull/9276#discussion_r43719970
--- Diff:
sql/core/src/main/scala/org/apache/spark/sql/execution/ExchangeCoordinator.scala
---
@@ -0,0 +1,260 @@
+/*
+ * 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
+
+import java.util.{Map => JMap, HashMap => JHashMap}
+
+import scala.collection.mutable.ArrayBuffer
+
+import org.apache.spark.{Logging, SimpleFutureAction, ShuffleDependency,
MapOutputStatistics}
+import org.apache.spark.rdd.RDD
+import org.apache.spark.sql.catalyst.InternalRow
+
+/**
+ * A coordinator used to determines how we shuffle data between stages
generated by Spark SQL.
+ * Right now, the work of this coordinator is to determine the number of
post-shuffle partitions
+ * for a stage that needs to fetch shuffle data from one or multiple
stages.
+ *
+ * A coordinator is constructed with three parameters, `numExchanges`,
+ * `targetPostShuffleInputSize`, and `minNumPostShufflePartitions`.
+ * - `numExchanges` is used to indicated that how many [[Exchange]]s that
will be registered to
+ * this coordinator. So, when we start to do any actual work, we have a
way to make sure that
+ * we have got expected number of [[Exchange]]s.
+ * - `targetPostShuffleInputSize` is the targeted size of a post-shuffle
partition's
+ * input data size. With this parameter, we can estimate the number of
post-shuffle partitions.
+ * This parameter is configured through
+ * `spark.sql.adaptive.shuffle.targetPostShuffleInputSize`.
+ * - `minNumPostShufflePartitions` is an optional parameter. If it is
defined, this coordinator
+ * will try to make sure that there are at least
`minNumPostShufflePartitions` post-shuffle
+ * partitions.
+ *
+ * The workflow of this coordinator is described as follows:
+ * - Before the execution of a [[SparkPlan]], for an [[Exchange]]
operator,
+ * if an [[ExchangeCoordinator]] is assigned to it, it registers itself
to this coordinator.
+ * This happens in the `doPrepare` method.
+ * - Once we start to execute a physical plan, an [[Exchange]] registered
to this coordinator will
+ * call `postShuffleRDD` to get its corresponding post-shuffle
[[ShuffledRowRDD]].
+ * If this coordinator has made the decision on how to shuffle data,
this [[Exchange]] will
+ * immediately get its corresponding post-shuffle [[ShuffledRowRDD]].
+ * - If this coordinator has not made the decision on how to shuffle
data, it will ask those
+ * registered [[Exchange]]s to submit their pre-shuffle stages. Then,
based on the the size
+ * statistics of pre-shuffle partitions, this coordinator will
determine the number of
+ * post-shuffle partitions and pack multiple pre-shuffle partitions
with continuous indices
+ * to a single post-shuffle partition whenever necessary.
+ * - Finally, this coordinator will create post-shuffle
[[ShuffledRowRDD]]s for all registered
+ * [[Exchange]]s. So, when an [[Exchange]] calls `postShuffleRDD`, this
coordinator can
+ * lookup the corresponding [[RDD]].
+ *
+ * The strategy used to determine the number of post-shuffle partitions is
described as follows.
+ * To determine the number of post-shuffle partitions, we have a target
input size for a
+ * post-shuffle partition. Once we have size statistics of pre-shuffle
partitions from stages
+ * corresponding to the registered [[Exchange]]s, we will do a pass of
those statistics and
+ * pack pre-shuffle partitions with continuous indices to a single
post-shuffle partition until
+ * the size of a post-shuffle partition is equal or greater than the
target size.
+ * For example, we have two stages with the following pre-shuffle
partition size statistics:
+ * stage 1: [100 MB, 20 MB, 100 MB, 10MB, 30 MB]
+ * stage 2: [10 MB, 10 MB, 70 MB, 5 MB, 5 MB]
+ * assuming the target input size is 128 MB, we will have three
post-shuffle partitions,
+ * which are:
+ * - post-shuffle partition 0: pre-shuffle partition 0 and 1
+ * - post-shuffle partition 1: pre-shuffle partition 2
+ * - post-shuffle partition 2: pre-shuffle partition 3 and 4
+ */
+private[sql] class ExchangeCoordinator(
+ numExchanges: Int,
+ advisoryTargetPostShuffleInputSize: Long,
+ minNumPostShufflePartitions: Option[Int] = None)
+ extends Logging {
+
+ // The registered Exchange operators.
+ private[this] val exchanges = ArrayBuffer[Exchange]()
+
+ // This map is used to lookup the post-shuffle ShuffledRowRDD for an
Exchange operator.
+ private[this] val postShuffleRDDs: JMap[Exchange, ShuffledRowRDD] =
+ new JHashMap[Exchange, ShuffledRowRDD](numExchanges)
+
+ // A boolean that indicates if this coordinator has made decision on how
to shuffle data.
+ // This variable will only be updated by doEstimationIfNecessary, which
is protected by
+ // synchronized.
+ @volatile private[this] var estimated: Boolean = false
+
+ /**
+ * Registers an [[Exchange]] operator to this coordinator. This method
is only allowed to be
+ * called in the `doPrepare` method of an [[Exchange]] operator.
+ */
+ def registerExchange(exchange: Exchange): Unit = synchronized {
+ exchanges += exchange
+ }
+
+ def isEstimated: Boolean = estimated
+
+ /**
+ * Estimates partition start indices for post-shuffle partitions based on
+ * mapOutputStatistics provided by all pre-shuffle stages.
+ */
+ private[sql] def estimatePartitionStartIndices(
+ mapOutputStatistics: Array[MapOutputStatistics]): Array[Int] = {
+ // If we have mapOutputStatistics.length <= numExchange, it is because
we do not submit
--- End diff --
`<` instead of `<=`?
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
