[SPARK-13523] [SQL] Reuse exchanges in a query

## What changes were proposed in this pull request?

It’s possible to have common parts in a query, for example, self join, it will be good to avoid the duplicated part to same CPUs and memory (Broadcast or cache).

Exchange will materialize the underlying RDD by shuffle or collect, it’s a great point to check duplicates and reuse them. Duplicated exchanges means they generate exactly the same result inside a query.

In order to find out the duplicated exchanges, we should be able to compare SparkPlan to check that they have same results or not. We already have that for LogicalPlan, so we should move that into QueryPlan to make it available for SparkPlan.

Once we can find the duplicated exchanges, we should replace all of them with same SparkPlan object (could be wrapped by ReusedExchage for explain), then the plan tree become a DAG. Since all the planner only work with tree, so this rule should be the last one for the entire planning.

After the rule, the plan will looks like:

```
WholeStageCodegen
:  +- Project [id#0L]
:     +- BroadcastHashJoin [id#0L], [id#2L], Inner, BuildRight, None
:        :- Project [id#0L]
:        :  +- BroadcastHashJoin [id#0L], [id#1L], Inner, BuildRight, None
:        :     :- Range 0, 1, 4, 1024, [id#0L]
:        :     +- INPUT
:        +- INPUT
:- BroadcastExchange HashedRelationBroadcastMode(true,List(id#1L),List(id#1L))
:  +- WholeStageCodegen
:     :  +- Range 0, 1, 4, 1024, [id#1L]
+- ReusedExchange [id#2L], BroadcastExchange HashedRelationBroadcastMode(true,List(id#1L),List(id#1L))
```

![bjoin](https://cloud.githubusercontent.com/assets/40902/13414787/209e8c5c-df0a-11e5-8a0f-edff69d89e83.png)

For three ways SortMergeJoin,
```
== Physical Plan ==
WholeStageCodegen
:  +- Project [id#0L]
:     +- SortMergeJoin [id#0L], [id#4L], None
:        :- INPUT
:        +- INPUT
:- WholeStageCodegen
:  :  +- Project [id#0L]
:  :     +- SortMergeJoin [id#0L], [id#3L], None
:  :        :- INPUT
:  :        +- INPUT
:  :- WholeStageCodegen
:  :  :  +- Sort [id#0L ASC], false, 0
:  :  :     +- INPUT
:  :  +- Exchange hashpartitioning(id#0L, 200), None
:  :     +- WholeStageCodegen
:  :        :  +- Range 0, 1, 4, 33554432, [id#0L]
:  +- WholeStageCodegen
:     :  +- Sort [id#3L ASC], false, 0
:     :     +- INPUT
:     +- ReusedExchange [id#3L], Exchange hashpartitioning(id#0L, 200), None
+- WholeStageCodegen
   :  +- Sort [id#4L ASC], false, 0
   :     +- INPUT
   +- ReusedExchange [id#4L], Exchange hashpartitioning(id#0L, 200), None
```
![sjoin](https://cloud.githubusercontent.com/assets/40902/13414790/27aea61c-df0a-11e5-8cbf-fbc985c31d95.png)

If the same ShuffleExchange or BroadcastExchange, execute()/executeBroadcast() will be called by different parents, they should cached the RDD/Broadcast, return the same one for all the parents.

## How was this patch tested?

Added some unit tests for this.  Had done some manual tests on TPCDS query Q59 and Q64, we can see some exchanges are re-used (this requires a change in PhysicalRDD to for sameResult, is be done in #11514 ).

Author: Davies Liu <[email protected]>

Closes #11403 from davies/dedup.

Author: Davies Liu

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/QueryPlan.scala

@@ -21,7 +21,7 @@ import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.catalyst.trees.TreeNode
 import org.apache.spark.sql.types.{DataType, StructType}
 
-abstract class QueryPlan[PlanType <: TreeNode[PlanType]] extends TreeNode[PlanType] {
+abstract class QueryPlan[PlanType <: QueryPlan[PlanType]] extends TreeNode[PlanType] {
   self: PlanType =>
 
   def output: Seq[Attribute]
@@ -237,4 +237,65 @@ abstract class QueryPlan[PlanType <: TreeNode[PlanType]] extends TreeNode[PlanTy
   }
 
   override def innerChildren: Seq[PlanType] = subqueries
+
+  /**
+   * Canonicalized copy of this query plan.
+   */
+  protected lazy val canonicalized: PlanType = this
+
+  /**
+   * Returns true when the given query plan will return the same results as this query plan.
+   *
+   * Since its likely undecidable to generally determine if two given plans will produce the same
+   * results, it is okay for this function to return false, even if the results are actually
+   * the same.  Such behavior will not affect correctness, only the application of performance
+   * enhancements like caching.  However, it is not acceptable to return true if the results could
+   * possibly be different.
+   *
+   * By default this function performs a modified version of equality that is tolerant of cosmetic
+   * differences like attribute naming and or expression id differences. Operators that
+   * can do better should override this function.
+   */
+  def sameResult(plan: PlanType): Boolean = {
+    val canonicalizedLeft = this.canonicalized
+    val canonicalizedRight = plan.canonicalized
+    canonicalizedLeft.getClass == canonicalizedRight.getClass &&
+      canonicalizedLeft.children.size == canonicalizedRight.children.size &&
+      canonicalizedLeft.cleanArgs == canonicalizedRight.cleanArgs &&
+      (canonicalizedLeft.children, canonicalizedRight.children).zipped.forall(_ sameResult _)
+  }
+
+  /**
+   * All the attributes that are used for this plan.
+   */
+  lazy val allAttributes: Seq[Attribute] = children.flatMap(_.output)
+
+  private def cleanExpression(e: Expression): Expression = e match {
+    case a: Alias =>
+      // As the root of the expression, Alias will always take an arbitrary exprId, we need
+      // to erase that for equality testing.
+      val cleanedExprId =
+        Alias(a.child, a.name)(ExprId(-1), a.qualifiers, isGenerated = a.isGenerated)
+      BindReferences.bindReference(cleanedExprId, allAttributes, allowFailures = true)
+    case other =>
+      BindReferences.bindReference(other, allAttributes, allowFailures = true)
+  }
+
+  /** Args that have cleaned such that differences in expression id should not affect equality */
+  protected lazy val cleanArgs: Seq[Any] = {
+    def cleanArg(arg: Any): Any = arg match {
+      case e: Expression => cleanExpression(e).canonicalized
+      case other => other
+    }
+
+    productIterator.map {
+      // Children are checked using sameResult above.
+      case tn: TreeNode[_] if containsChild(tn) => null
+      case e: Expression => cleanArg(e)
+      case s: Option[_] => s.map(cleanArg)
+      case s: Seq[_] => s.map(cleanArg)
+      case m: Map[_, _] => m.mapValues(cleanArg)
+      case other => other
+    }.toSeq
+  }
 }

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/LogicalPlan.scala

@@ -114,60 +114,7 @@ abstract class LogicalPlan extends QueryPlan[LogicalPlan] with Logging {
    */
   def childrenResolved: Boolean = children.forall(_.resolved)
 
-  /**
-   * Returns true when the given logical plan will return the same results as this logical plan.
-   *
-   * Since its likely undecidable to generally determine if two given plans will produce the same
-   * results, it is okay for this function to return false, even if the results are actually
-   * the same.  Such behavior will not affect correctness, only the application of performance
-   * enhancements like caching.  However, it is not acceptable to return true if the results could
-   * possibly be different.
-   *
-   * By default this function performs a modified version of equality that is tolerant of cosmetic
-   * differences like attribute naming and or expression id differences.  Logical operators that
-   * can do better should override this function.
-   */
-  def sameResult(plan: LogicalPlan): Boolean = {
-    val cleanLeft = EliminateSubqueryAliases(this)
-    val cleanRight = EliminateSubqueryAliases(plan)
-
-    cleanLeft.getClass == cleanRight.getClass &&
-      cleanLeft.children.size == cleanRight.children.size && {
-      logDebug(
-        s"[${cleanRight.cleanArgs.mkString(", ")}] == [${cleanLeft.cleanArgs.mkString(", ")}]")
-      cleanRight.cleanArgs == cleanLeft.cleanArgs
-    } &&
-    (cleanLeft.children, cleanRight.children).zipped.forall(_ sameResult _)
-  }
-
-  /** Args that have cleaned such that differences in expression id should not affect equality */
-  protected lazy val cleanArgs: Seq[Any] = {
-    val input = children.flatMap(_.output)
-    def cleanExpression(e: Expression) = e match {
-      case a: Alias =>
-        // As the root of the expression, Alias will always take an arbitrary exprId, we need
-        // to erase that for equality testing.
-        val cleanedExprId =
-          Alias(a.child, a.name)(ExprId(-1), a.qualifiers, isGenerated = a.isGenerated)
-        BindReferences.bindReference(cleanedExprId, input, allowFailures = true)
-      case other => BindReferences.bindReference(other, input, allowFailures = true)
-    }
-
-    productIterator.map {
-      // Children are checked using sameResult above.
-      case tn: TreeNode[_] if containsChild(tn) => null
-      case e: Expression => cleanExpression(e)
-      case s: Option[_] => s.map {
-        case e: Expression => cleanExpression(e)
-        case other => other
-      }
-      case s: Seq[_] => s.map {
-        case e: Expression => cleanExpression(e)
-        case other => other
-      }
-      case other => other
-    }.toSeq
-  }
+  override lazy val canonicalized: LogicalPlan = EliminateSubqueryAliases(this)
 
   /**
    * Optionally resolves the given strings to a [[NamedExpression]] using the input from all child

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/physical/broadcastMode.scala

@@ -25,6 +25,11 @@ import org.apache.spark.sql.catalyst.InternalRow
  */
 trait BroadcastMode {
   def transform(rows: Array[InternalRow]): Any
+
+  /**
+   * Returns true iff this [[BroadcastMode]] generates the same result as `other`.
+   */
+  def compatibleWith(other: BroadcastMode): Boolean
 }
 
 /**
@@ -33,4 +38,8 @@ trait BroadcastMode {
 case object IdentityBroadcastMode extends BroadcastMode {
   // TODO: pack the UnsafeRows into single bytes array.
   override def transform(rows: Array[InternalRow]): Array[InternalRow] = rows
+
+  override def compatibleWith(other: BroadcastMode): Boolean = {
+    this eq other
+  }
 }

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

@@ -18,6 +18,7 @@
 package org.apache.spark.sql.execution
 
 import org.apache.spark.annotation.DeveloperApi
+import org.apache.spark.sql.execution.exchange.ReusedExchange
 import org.apache.spark.sql.execution.metric.SQLMetricInfo
 import org.apache.spark.util.Utils
 
@@ -31,13 +32,28 @@ class SparkPlanInfo(
     val simpleString: String,
     val children: Seq[SparkPlanInfo],
     val metadata: Map[String, String],
-    val metrics: Seq[SQLMetricInfo])
+    val metrics: Seq[SQLMetricInfo]) {
+
+  override def hashCode(): Int = {
+    // hashCode of simpleString should be good enough to distinguish the plans from each other
+    // within a plan
+    simpleString.hashCode
+  }
+
+  override def equals(other: Any): Boolean = other match {
+    case o: SparkPlanInfo =>
+      nodeName == o.nodeName && simpleString == o.simpleString && children == o.children
+    case _ => false
+  }
+}
 
 private[sql] object SparkPlanInfo {
 
   def fromSparkPlan(plan: SparkPlan): SparkPlanInfo = {
-
-    val children = plan.children ++ plan.subqueries
+    val children = plan match {
+      case ReusedExchange(_, child) => child :: Nil
+      case _ => plan.children ++ plan.subqueries
+    }
     val metrics = plan.metrics.toSeq.map { case (key, metric) =>
       new SQLMetricInfo(metric.name.getOrElse(key), metric.id,
         Utils.getFormattedClassName(metric.param))

sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/TungstenAggregate.scala

@@ -46,6 +46,10 @@ case class TungstenAggregate(
 
   require(TungstenAggregate.supportsAggregate(aggregateBufferAttributes))
 
+  override lazy val allAttributes: Seq[Attribute] =
+    child.output ++ aggregateBufferAttributes ++ aggregateAttributes ++
+      aggregateExpressions.flatMap(_.aggregateFunction.inputAggBufferAttributes)
+
   override private[sql] lazy val metrics = Map(
     "numOutputRows" -> SQLMetrics.createLongMetric(sparkContext, "number of output rows"),
     "dataSize" -> SQLMetrics.createSizeMetric(sparkContext, "data size"),

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

@@ -166,6 +166,9 @@ case class Range(
   private[sql] override lazy val metrics = Map(
     "numOutputRows" -> SQLMetrics.createLongMetric(sparkContext, "number of output rows"))
 
+  // output attributes should not affect the results
+  override lazy val cleanArgs: Seq[Any] = Seq(start, step, numSlices, numElements)
+
   override def upstreams(): Seq[RDD[InternalRow]] = {
     sqlContext.sparkContext.parallelize(0 until numSlices, numSlices)
       .map(i => InternalRow(i)) :: Nil

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

@@ -34,12 +34,16 @@ import org.apache.spark.util.ThreadUtils
  */
 case class BroadcastExchange(
     mode: BroadcastMode,
-    child: SparkPlan) extends UnaryNode {
-
-  override def output: Seq[Attribute] = child.output
+    child: SparkPlan) extends Exchange {
 
   override def outputPartitioning: Partitioning = BroadcastPartitioning(mode)
 
+  override def sameResult(plan: SparkPlan): Boolean = plan match {
+    case p: BroadcastExchange =>
+      mode.compatibleWith(p.mode) && child.sameResult(p.child)
+    case _ => false
+  }
+
   @transient
   private val timeout: Duration = {
     val timeoutValue = sqlContext.conf.broadcastTimeout

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

@@ -0,0 +1,92 @@
+/*
+ * 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.exchange
+
+import scala.collection.mutable
+import scala.collection.mutable.ArrayBuffer
+
+import org.apache.spark.broadcast
+import org.apache.spark.rdd.RDD
+import org.apache.spark.sql.SQLContext
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.catalyst.expressions.Attribute
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.{LeafNode, SparkPlan, UnaryNode}
+import org.apache.spark.sql.types.StructType
+
+/**
+ * An interface for exchanges.
+ */
+abstract class Exchange extends UnaryNode {
+  override def output: Seq[Attribute] = child.output
+}
+
+/**
+ * A wrapper for reused exchange to have different output, because two exchanges which produce
+ * logically identical output will have distinct sets of output attribute ids, so we need to
+ * preserve the original ids because they're what downstream operators are expecting.
+ */
+case class ReusedExchange(override val output: Seq[Attribute], child: Exchange) extends LeafNode {
+
+  override def sameResult(plan: SparkPlan): Boolean = {
+    // Ignore this wrapper. `plan` could also be a ReusedExchange, so we reverse the order here.
+    plan.sameResult(child)
+  }
+
+  def doExecute(): RDD[InternalRow] = {
+    child.execute()
+  }
+
+  override protected[sql] def doExecuteBroadcast[T](): broadcast.Broadcast[T] = {
+    child.executeBroadcast()
+  }
+
+  // Do not repeat the same tree in explain.
+  override def treeChildren: Seq[SparkPlan] = Nil
+}
+
+/**
+ * Find out duplicated exchanges in the spark plan, then use the same exchange for all the
+ * references.
+ */
+private[sql] case class ReuseExchange(sqlContext: SQLContext) extends Rule[SparkPlan] {
+
+  def apply(plan: SparkPlan): SparkPlan = {
+    if (!sqlContext.conf.exchangeReuseEnabled) {
+      return plan
+    }
+    // Build a hash map using schema of exchanges to avoid O(N*N) sameResult calls.
+    val exchanges = mutable.HashMap[StructType, ArrayBuffer[Exchange]]()
+    plan.transformUp {
+      case exchange: Exchange =>
+        // the exchanges that have same results usually also have same schemas (same column names).
+        val sameSchema = exchanges.getOrElseUpdate(exchange.schema, ArrayBuffer[Exchange]())
+        val samePlan = sameSchema.find { e =>
+          exchange.sameResult(e)
+        }
+        if (samePlan.isDefined) {
+          // Keep the output of this exchange, the following plans require that to resolve
+          // attributes.
+          ReusedExchange(exchange.output, samePlan.get)
+        } else {
+          sameSchema += exchange
+          exchange
+        }
+    }
+  }
+}

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

@@ -38,7 +38,7 @@ import org.apache.spark.util.MutablePair
 case class ShuffleExchange(
     var newPartitioning: Partitioning,
     child: SparkPlan,
-    @transient coordinator: Option[ExchangeCoordinator]) extends UnaryNode {
+    @transient coordinator: Option[ExchangeCoordinator]) extends Exchange {
 
   override def nodeName: String = {
     val extraInfo = coordinator match {
@@ -55,8 +55,6 @@ case class ShuffleExchange(
 
   override def outputPartitioning: Partitioning = newPartitioning
 
-  override def output: Seq[Attribute] = child.output
-
   private val serializer: Serializer = new UnsafeRowSerializer(child.output.size)
 
   override protected def doPrepare(): Unit = {
@@ -103,16 +101,25 @@ case class ShuffleExchange(
     new ShuffledRowRDD(shuffleDependency, specifiedPartitionStartIndices)
   }
 
+  /**
+   * Caches the created ShuffleRowRDD so we can reuse that.
+   */
+  private var cachedShuffleRDD: ShuffledRowRDD = null
+
   protected override def doExecute(): RDD[InternalRow] = attachTree(this, "execute") {
-    coordinator match {
-      case Some(exchangeCoordinator) =>
-        val shuffleRDD = exchangeCoordinator.postShuffleRDD(this)
-        assert(shuffleRDD.partitions.length == newPartitioning.numPartitions)
-        shuffleRDD
-      case None =>
-        val shuffleDependency = prepareShuffleDependency()
-        preparePostShuffleRDD(shuffleDependency)
+    // Returns the same ShuffleRowRDD if this plan is used by multiple plans.
+    if (cachedShuffleRDD == null) {
+      cachedShuffleRDD = coordinator match {
+        case Some(exchangeCoordinator) =>
+          val shuffleRDD = exchangeCoordinator.postShuffleRDD(this)
+          assert(shuffleRDD.partitions.length == newPartitioning.numPartitions)
+          shuffleRDD
+        case None =>
+          val shuffleDependency = prepareShuffleDependency()
+          preparePostShuffleRDD(shuffleDependency)
+      }
     }
+    cachedShuffleRDD
   }
 }