refine

Author: cloud-fan

sql/core/src/main/scala/org/apache/spark/sql/execution/SparkStrategies.scala

@@ -32,7 +32,7 @@ import org.apache.spark.sql.catalyst.streaming.{InternalOutputModes, StreamingRe
 import org.apache.spark.sql.execution.aggregate.AggUtils
 import org.apache.spark.sql.execution.columnar.{InMemoryRelation, InMemoryTableScanExec}
 import org.apache.spark.sql.execution.command._
-import org.apache.spark.sql.execution.exchange.{ShuffleExchangeExec, ShuffleOrigin}
+import org.apache.spark.sql.execution.exchange.{REPARTITION, REPARTITION_WITH_NUM, ShuffleExchangeExec}
 import org.apache.spark.sql.execution.python._
 import org.apache.spark.sql.execution.streaming._
 import org.apache.spark.sql.execution.streaming.sources.MemoryPlan
@@ -670,7 +670,7 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
       case logical.Repartition(numPartitions, shuffle, child) =>
         if (shuffle) {
           ShuffleExchangeExec(RoundRobinPartitioning(numPartitions),
-            planLater(child), ShuffleOrigin.REPARTITION_WITH_NUM) :: Nil
+            planLater(child), REPARTITION_WITH_NUM) :: Nil
         } else {
           execution.CoalesceExec(numPartitions, planLater(child)) :: Nil
         }
@@ -704,9 +704,9 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
         execution.RangeExec(r) :: Nil
       case r: logical.RepartitionByExpression =>
         val shuffleOrigin = if (r.optNumPartitions.isEmpty) {
-          ShuffleOrigin.REPARTITION
+          REPARTITION
         } else {
-          ShuffleOrigin.REPARTITION_WITH_NUM
+          REPARTITION_WITH_NUM
         }
         exchange.ShuffleExchangeExec(r.partitioning, planLater(r.child), shuffleOrigin) :: Nil
       case ExternalRDD(outputObjAttr, rdd) => ExternalRDDScanExec(outputObjAttr, rdd) :: Nil

sql/core/src/main/scala/org/apache/spark/sql/execution/adaptive/CoalesceShufflePartitions.scala

@@ -18,8 +18,10 @@
 package org.apache.spark.sql.execution.adaptive
 
 import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.plans.physical.SinglePartition
 import org.apache.spark.sql.catalyst.rules.Rule
 import org.apache.spark.sql.execution.SparkPlan
+import org.apache.spark.sql.execution.exchange.{ENSURE_REQUIREMENTS, REPARTITION, ShuffleExchangeLike}
 import org.apache.spark.sql.internal.SQLConf
 
 /**
@@ -47,7 +49,7 @@ case class CoalesceShufflePartitions(session: SparkSession) extends Rule[SparkPl
     val shuffleStages = collectShuffleStages(plan)
     // ShuffleExchanges introduced by repartition do not support changing the number of partitions.
     // We change the number of partitions in the stage only if all the ShuffleExchanges support it.
-    if (!shuffleStages.forall(_.shuffle.canChangeNumPartitions)) {
+    if (!shuffleStages.forall(s => supportCoalesce(s.shuffle))) {
       plan
     } else {
       // `ShuffleQueryStageExec#mapStats` returns None when the input RDD has 0 partitions,
@@ -82,4 +84,9 @@ case class CoalesceShufflePartitions(session: SparkSession) extends Rule[SparkPl
       }
     }
   }
+
+  private def supportCoalesce(s: ShuffleExchangeLike): Boolean = {
+    s.outputPartitioning != SinglePartition &&
+      (s.shuffleOrigin == ENSURE_REQUIREMENTS || s.shuffleOrigin == REPARTITION)
+  }
 }

sql/core/src/main/scala/org/apache/spark/sql/execution/adaptive/OptimizeLocalShuffleReader.scala

@@ -18,9 +18,10 @@
 package org.apache.spark.sql.execution.adaptive
 
 import org.apache.spark.sql.catalyst.optimizer.{BuildLeft, BuildRight, BuildSide}
+import org.apache.spark.sql.catalyst.plans.physical.SinglePartition
 import org.apache.spark.sql.catalyst.rules.Rule
 import org.apache.spark.sql.execution._
-import org.apache.spark.sql.execution.exchange.{EnsureRequirements, ShuffleExchangeExec}
+import org.apache.spark.sql.execution.exchange.{ENSURE_REQUIREMENTS, EnsureRequirements, ShuffleExchangeExec, ShuffleExchangeLike}
 import org.apache.spark.sql.execution.joins.BroadcastHashJoinExec
 import org.apache.spark.sql.internal.SQLConf
 
@@ -136,9 +137,13 @@ object OptimizeLocalShuffleReader extends Rule[SparkPlan] {
 
   def canUseLocalShuffleReader(plan: SparkPlan): Boolean = plan match {
     case s: ShuffleQueryStageExec =>
-      s.shuffle.canChangePartitioning && s.mapStats.isDefined
+      s.mapStats.isDefined && supportLocalReader(s.shuffle)
     case CustomShuffleReaderExec(s: ShuffleQueryStageExec, partitionSpecs) =>
-      s.shuffle.canChangePartitioning && s.mapStats.isDefined && partitionSpecs.nonEmpty
+      s.mapStats.isDefined && partitionSpecs.nonEmpty && supportLocalReader(s.shuffle)
     case _ => false
   }
+
+  private def supportLocalReader(s: ShuffleExchangeLike): Boolean = {
+    s.outputPartitioning != SinglePartition && s.shuffleOrigin == ENSURE_REQUIREMENTS
+  }
 }

sql/core/src/main/scala/org/apache/spark/sql/execution/exchange/ShuffleExchangeExec.scala

@@ -56,25 +56,10 @@ trait ShuffleExchangeLike extends Exchange {
    */
   def numPartitions: Int
 
-  def shuffleOrigin: ShuffleOrigin.Value
-
-  /**
-   * Returns whether the shuffle partition number can be changed.
-   */
-  final def canChangeNumPartitions: Boolean = {
-    // If users specify the num partitions via APIs like `repartition(5, col)`, we shouldn't change
-    // it. For `SinglePartition`, it requires exactly one partition and we can't change it either.
-    shuffleOrigin != ShuffleOrigin.REPARTITION_WITH_NUM && outputPartitioning != SinglePartition
-  }
-
   /**
-   * Returns whether the shuffle output data partitioning can be changed.
+   * The origin of this shuffle operator.
    */
-  final def canChangePartitioning: Boolean = {
-    // If users specify the partitioning via APIs like `repartition(col)`, we shouldn't change it.
-    // For `SinglePartition`, itself is a special partitioning and we can't change it either.
-    shuffleOrigin == ShuffleOrigin.ENSURE_REQUIREMENTS && outputPartitioning != SinglePartition
-  }
+  def shuffleOrigin: ShuffleOrigin
 
   /**
    * The asynchronous job that materializes the shuffle.
@@ -93,27 +78,28 @@ trait ShuffleExchangeLike extends Exchange {
 }
 
 // Describes where the shuffle operator comes from.
-object ShuffleOrigin extends Enumeration {
-  type ShuffleOrigin = Value
-  // Indicates that the shuffle operator was added by the internal `EnsureRequirements` rule. It
-  // means that the shuffle operator is used to ensure internal data partitioning requirements and
-  // Spark is free to optimize it as long as the requirements are still ensured.
-  val ENSURE_REQUIREMENTS = Value
-  // Indicates that the shuffle operator was added by the user-specified repartition operator. Spark
-  // can still optimize it via changing shuffle partition number, as data partitioning won't change.
-  val REPARTITION = Value
-  // Indicates that the shuffle operator was added by the user-specified repartition operator with
-  // a certain partition number. Spark can't optimize it.
-  val REPARTITION_WITH_NUM = Value
-}
+sealed trait ShuffleOrigin
+
+// Indicates that the shuffle operator was added by the internal `EnsureRequirements` rule. It
+// means that the shuffle operator is used to ensure internal data partitioning requirements and
+// Spark is free to optimize it as long as the requirements are still ensured.
+case object ENSURE_REQUIREMENTS extends ShuffleOrigin
+
+// Indicates that the shuffle operator was added by the user-specified repartition operator. Spark
+// can still optimize it via changing shuffle partition number, as data partitioning won't change.
+case object REPARTITION extends ShuffleOrigin
+
+// Indicates that the shuffle operator was added by the user-specified repartition operator with
+// a certain partition number. Spark can't optimize it.
+case object REPARTITION_WITH_NUM extends ShuffleOrigin
 
 /**
  * Performs a shuffle that will result in the desired partitioning.
  */
 case class ShuffleExchangeExec(
     override val outputPartitioning: Partitioning,
     child: SparkPlan,
-    shuffleOrigin: ShuffleOrigin.Value = ShuffleOrigin.ENSURE_REQUIREMENTS)
+    shuffleOrigin: ShuffleOrigin = ENSURE_REQUIREMENTS)
   extends ShuffleExchangeLike {
 
   private lazy val writeMetrics =

sql/core/src/test/resources/sql-tests/results/explain-aqe.sql.out

@@ -67,10 +67,10 @@ Aggregate [sum(distinct cast(val#x as bigint)) AS sum(DISTINCT val)#xL]
 == Physical Plan ==
 AdaptiveSparkPlan isFinalPlan=false
 +- HashAggregate(keys=[], functions=[sum(distinct cast(val#x as bigint)#xL)], output=[sum(DISTINCT val)#xL])
-   +- Exchange SinglePartition, UNSPECIFIED, [id=#x]
+   +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
       +- HashAggregate(keys=[], functions=[partial_sum(distinct cast(val#x as bigint)#xL)], output=[sum#xL])
          +- HashAggregate(keys=[cast(val#x as bigint)#xL], functions=[], output=[cast(val#x as bigint)#xL])
-            +- Exchange hashpartitioning(cast(val#x as bigint)#xL, 4), UNSPECIFIED, [id=#x]
+            +- Exchange hashpartitioning(cast(val#x as bigint)#xL, 4), ENSURE_REQUIREMENTS, [id=#x]
                +- HashAggregate(keys=[cast(val#x as bigint) AS cast(val#x as bigint)#xL], functions=[], output=[cast(val#x as bigint)#xL])
                   +- FileScan parquet default.explain_temp1[val#x] Batched: true, DataFilters: [], Format: Parquet, Location [not included in comparison]/{warehouse_dir}/explain_temp1], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<val:int>
 
@@ -116,7 +116,7 @@ Results [2]: [key#x, max#x]
 
 (4) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (5) HashAggregate
 Input [2]: [key#x, max#x]
@@ -127,7 +127,7 @@ Results [2]: [key#x, max(val#x)#x AS max(val)#x]
 
 (6) Exchange
 Input [2]: [key#x, max(val)#x]
-Arguments: rangepartitioning(key#x ASC NULLS FIRST, 4), UNSPECIFIED, [id=#x]
+Arguments: rangepartitioning(key#x ASC NULLS FIRST, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (7) Sort
 Input [2]: [key#x, max(val)#x]
@@ -179,7 +179,7 @@ Results [2]: [key#x, max#x]
 
 (4) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (5) HashAggregate
 Input [2]: [key#x, max#x]
@@ -254,7 +254,7 @@ Results [2]: [key#x, val#x]
 
 (7) Exchange
 Input [2]: [key#x, val#x]
-Arguments: hashpartitioning(key#x, val#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, val#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (8) HashAggregate
 Input [2]: [key#x, val#x]
@@ -576,7 +576,7 @@ Results [2]: [key#x, max#x]
 
 (4) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (5) HashAggregate
 Input [2]: [key#x, max#x]
@@ -605,7 +605,7 @@ Results [2]: [key#x, max#x]
 
 (9) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (10) HashAggregate
 Input [2]: [key#x, max#x]
@@ -687,7 +687,7 @@ Results [3]: [count#xL, sum#xL, count#xL]
 
 (3) Exchange
 Input [3]: [count#xL, sum#xL, count#xL]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (4) HashAggregate
 Input [3]: [count#xL, sum#xL, count#xL]
@@ -732,7 +732,7 @@ Results [2]: [key#x, buf#x]
 
 (3) Exchange
 Input [2]: [key#x, buf#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (4) ObjectHashAggregate
 Input [2]: [key#x, buf#x]
@@ -783,7 +783,7 @@ Results [2]: [key#x, min#x]
 
 (4) Exchange
 Input [2]: [key#x, min#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (5) Sort
 Input [2]: [key#x, min#x]

sql/core/src/test/resources/sql-tests/results/explain.sql.out

@@ -66,10 +66,10 @@ Aggregate [sum(distinct cast(val#x as bigint)) AS sum(DISTINCT val)#xL]
 
 == Physical Plan ==
 *HashAggregate(keys=[], functions=[sum(distinct cast(val#x as bigint)#xL)], output=[sum(DISTINCT val)#xL])
-+- Exchange SinglePartition, UNSPECIFIED, [id=#x]
++- Exchange SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
    +- *HashAggregate(keys=[], functions=[partial_sum(distinct cast(val#x as bigint)#xL)], output=[sum#xL])
       +- *HashAggregate(keys=[cast(val#x as bigint)#xL], functions=[], output=[cast(val#x as bigint)#xL])
-         +- Exchange hashpartitioning(cast(val#x as bigint)#xL, 4), UNSPECIFIED, [id=#x]
+         +- Exchange hashpartitioning(cast(val#x as bigint)#xL, 4), ENSURE_REQUIREMENTS, [id=#x]
             +- *HashAggregate(keys=[cast(val#x as bigint) AS cast(val#x as bigint)#xL], functions=[], output=[cast(val#x as bigint)#xL])
                +- *ColumnarToRow
                   +- FileScan parquet default.explain_temp1[val#x] Batched: true, DataFilters: [], Format: Parquet, Location [not included in comparison]/{warehouse_dir}/explain_temp1], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<val:int>
@@ -119,7 +119,7 @@ Results [2]: [key#x, max#x]
 
 (5) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (6) HashAggregate [codegen id : 2]
 Input [2]: [key#x, max#x]
@@ -130,7 +130,7 @@ Results [2]: [key#x, max(val#x)#x AS max(val)#x]
 
 (7) Exchange
 Input [2]: [key#x, max(val)#x]
-Arguments: rangepartitioning(key#x ASC NULLS FIRST, 4), UNSPECIFIED, [id=#x]
+Arguments: rangepartitioning(key#x ASC NULLS FIRST, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (8) Sort [codegen id : 3]
 Input [2]: [key#x, max(val)#x]
@@ -181,7 +181,7 @@ Results [2]: [key#x, max#x]
 
 (5) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (6) HashAggregate [codegen id : 2]
 Input [2]: [key#x, max#x]
@@ -259,7 +259,7 @@ Results [2]: [key#x, val#x]
 
 (9) Exchange
 Input [2]: [key#x, val#x]
-Arguments: hashpartitioning(key#x, val#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, val#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (10) HashAggregate [codegen id : 4]
 Input [2]: [key#x, val#x]
@@ -452,7 +452,7 @@ Results [1]: [max#x]
 
 (9) Exchange
 Input [1]: [max#x]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (10) HashAggregate [codegen id : 2]
 Input [1]: [max#x]
@@ -498,7 +498,7 @@ Results [1]: [max#x]
 
 (16) Exchange
 Input [1]: [max#x]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (17) HashAggregate [codegen id : 2]
 Input [1]: [max#x]
@@ -580,7 +580,7 @@ Results [1]: [max#x]
 
 (9) Exchange
 Input [1]: [max#x]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (10) HashAggregate [codegen id : 2]
 Input [1]: [max#x]
@@ -626,7 +626,7 @@ Results [2]: [sum#x, count#xL]
 
 (16) Exchange
 Input [2]: [sum#x, count#xL]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (17) HashAggregate [codegen id : 2]
 Input [2]: [sum#x, count#xL]
@@ -690,7 +690,7 @@ Results [2]: [sum#x, count#xL]
 
 (7) Exchange
 Input [2]: [sum#x, count#xL]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (8) HashAggregate [codegen id : 2]
 Input [2]: [sum#x, count#xL]
@@ -810,7 +810,7 @@ Results [2]: [key#x, max#x]
 
 (5) Exchange
 Input [2]: [key#x, max#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (6) HashAggregate [codegen id : 4]
 Input [2]: [key#x, max#x]
@@ -901,7 +901,7 @@ Results [3]: [count#xL, sum#xL, count#xL]
 
 (4) Exchange
 Input [3]: [count#xL, sum#xL, count#xL]
-Arguments: SinglePartition, UNSPECIFIED, [id=#x]
+Arguments: SinglePartition, ENSURE_REQUIREMENTS, [id=#x]
 
 (5) HashAggregate [codegen id : 2]
 Input [3]: [count#xL, sum#xL, count#xL]
@@ -945,7 +945,7 @@ Results [2]: [key#x, buf#x]
 
 (4) Exchange
 Input [2]: [key#x, buf#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (5) ObjectHashAggregate
 Input [2]: [key#x, buf#x]
@@ -995,7 +995,7 @@ Results [2]: [key#x, min#x]
 
 (5) Exchange
 Input [2]: [key#x, min#x]
-Arguments: hashpartitioning(key#x, 4), UNSPECIFIED, [id=#x]
+Arguments: hashpartitioning(key#x, 4), ENSURE_REQUIREMENTS, [id=#x]
 
 (6) Sort [codegen id : 2]
 Input [2]: [key#x, min#x]

sql/core/src/test/scala/org/apache/spark/sql/SparkSessionExtensionSuite.scala

@@ -766,7 +766,7 @@ case class PreRuleReplaceAddWithBrokenVersion() extends Rule[SparkPlan] {
 case class MyShuffleExchangeExec(delegate: ShuffleExchangeExec) extends ShuffleExchangeLike {
   override def numMappers: Int = delegate.numMappers
   override def numPartitions: Int = delegate.numPartitions
-  override def shuffleOrigin: ShuffleOrigin.Value = {
+  override def shuffleOrigin: ShuffleOrigin = {
     delegate.shuffleOrigin
   }
   override def mapOutputStatisticsFuture: Future[MapOutputStatistics] =