[SPARK-9729] [SPARK-9363] [SQL] Use sort merge join for left and right outer join

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/JoinedRow.scala

@@ -37,20 +37,20 @@ class JoinedRow extends InternalRow {
   }
 
   /** Updates this JoinedRow to used point at two new base rows.  Returns itself. */
-  def apply(r1: InternalRow, r2: InternalRow): InternalRow = {
+  def apply(r1: InternalRow, r2: InternalRow): JoinedRow = {
     row1 = r1
     row2 = r2
     this
   }
 
   /** Updates this JoinedRow by updating its left base row.  Returns itself. */
-  def withLeft(newLeft: InternalRow): InternalRow = {
+  def withLeft(newLeft: InternalRow): JoinedRow = {
     row1 = newLeft
     this
   }
 
   /** Updates this JoinedRow by updating its right base row.  Returns itself. */
-  def withRight(newRight: InternalRow): InternalRow = {
+  def withRight(newRight: InternalRow): JoinedRow = {
     row2 = newRight
     this
   }

sql/core/src/main/scala/org/apache/spark/sql/SQLContext.scala

@@ -873,7 +873,7 @@ class SQLContext(@transient val sparkContext: SparkContext)
       HashAggregation ::
       Aggregation ::
       LeftSemiJoin ::
-      HashJoin ::
+      EquiJoinSelection ::
       InMemoryScans ::
       BasicOperators ::
       CartesianProduct ::

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

@@ -0,0 +1,93 @@
+/*
+ * 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.NoSuchElementException
+
+import org.apache.spark.sql.catalyst.InternalRow
+
+/**
+ * An internal iterator interface which presents a more restrictive API than
+ * [[scala.collection.Iterator]].
+ *
+ * One major departure from the Scala iterator API is the fusing of the `hasNext()` and `next()`
+ * calls: Scala's iterator allows users to call `hasNext()` without immediately advancing the
+ * iterator to consume the next row, whereas RowIterator combines these calls into a single
+ * [[advanceNext()]] method.
+ */
+private[sql] abstract class RowIterator {
+  /**
+   * Advance this iterator by a single row. Returns `false` if this iterator has no more rows
+   * and `true` otherwise. If this returns `true`, then the new row can be retrieved by calling
+   * [[getRow]].
+   */
+  def advanceNext(): Boolean
+
+  /**
+   * Retrieve the row from this iterator. This method is idempotent. It is illegal to call this
+   * method after [[advanceNext()]] has returned `false`.
+   */
+  def getRow: InternalRow
+
+  /**
+   * Convert this RowIterator into a [[scala.collection.Iterator]].
+   */
+  def toScala: Iterator[InternalRow] = new RowIteratorToScala(this)
+}
+
+object RowIterator {
+  def fromScala(scalaIter: Iterator[InternalRow]): RowIterator = {
+    scalaIter match {
+      case wrappedRowIter: RowIteratorToScala => wrappedRowIter.rowIter
+      case _ => new RowIteratorFromScala(scalaIter)
+    }
+  }
+}
+
+private final class RowIteratorToScala(val rowIter: RowIterator) extends Iterator[InternalRow] {
+  private [this] var hasNextWasCalled: Boolean = false
+  private [this] var _hasNext: Boolean = false
+  override def hasNext: Boolean = {
+    // Idempotency:
+    if (!hasNextWasCalled) {
+      _hasNext = rowIter.advanceNext()
+      hasNextWasCalled = true
+    }
+    _hasNext
+  }
+  override def next(): InternalRow = {
+    if (!hasNext) throw new NoSuchElementException
+    hasNextWasCalled = false
+    rowIter.getRow
+  }
+}
+
+private final class RowIteratorFromScala(scalaIter: Iterator[InternalRow]) extends RowIterator {
+  private[this] var _next: InternalRow = null
+  override def advanceNext(): Boolean = {
+    if (scalaIter.hasNext) {
+      _next = scalaIter.next()
+      true
+    } else {
+      _next = null
+      false
+    }
+  }
+  override def getRow: InternalRow = _next
+  override def toScala: Iterator[InternalRow] = scalaIter
+}

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

@@ -63,19 +63,23 @@ private[sql] abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
   }
 
   /**
-   * Uses the ExtractEquiJoinKeys pattern to find joins where at least some of the predicates can be
-   * evaluated by matching hash keys.
+   * Uses the [[ExtractEquiJoinKeys]] pattern to find joins where at least some of the predicates
+   * can be evaluated by matching join keys.
    *
-   * This strategy applies a simple optimization based on the estimates of the physical sizes of
-   * the two join sides.  When planning a [[joins.BroadcastHashJoin]], if one side has an
-   * estimated physical size smaller than the user-settable threshold
-   * [[org.apache.spark.sql.SQLConf.AUTO_BROADCASTJOIN_THRESHOLD]], the planner would mark it as the
-   * ''build'' relation and mark the other relation as the ''stream'' side.  The build table will be
-   * ''broadcasted'' to all of the executors involved in the join, as a
-   * [[org.apache.spark.broadcast.Broadcast]] object.  If both estimates exceed the threshold, they
-   * will instead be used to decide the build side in a [[joins.ShuffledHashJoin]].
+   * Join implementations are chosen with the following precedence:
+   *
+   * - Broadcast: if one side of the join has an estimated physical size that is smaller than the
+   *     user-configurable [[org.apache.spark.sql.SQLConf.AUTO_BROADCASTJOIN_THRESHOLD]] threshold
+   *     or if that side has an explicit broadcast hint (e.g. the user applied the
+   *     [[org.apache.spark.sql.functions.broadcast()]] function to a DataFrame), then that side
+   *     of the join will be broadcasted and the other side will be streamed, with no shuffling
+   *     performed. If both sides of the join are eligible to be broadcasted then the
+   * - Sort merge: if the matching join keys are sortable and
+   *     [[org.apache.spark.sql.SQLConf.SORTMERGE_JOIN]] is enabled (default), then sort merge join
+   *     will be used.
+   * - Hash: will be chosen if neither of the above optimizations apply to this join.
    */
-  object HashJoin extends Strategy with PredicateHelper {
+  object EquiJoinSelection extends Strategy with PredicateHelper {
 
     private[this] def makeBroadcastHashJoin(
         leftKeys: Seq[Expression],
@@ -90,14 +94,15 @@ private[sql] abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
     }
 
     def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
+
+      // --- Inner joins --------------------------------------------------------------------------
+
       case ExtractEquiJoinKeys(Inner, leftKeys, rightKeys, condition, left, CanBroadcast(right)) =>
         makeBroadcastHashJoin(leftKeys, rightKeys, left, right, condition, joins.BuildRight)
 
       case ExtractEquiJoinKeys(Inner, leftKeys, rightKeys, condition, CanBroadcast(left), right) =>
         makeBroadcastHashJoin(leftKeys, rightKeys, left, right, condition, joins.BuildLeft)
 
-      // If the sort merge join option is set, we want to use sort merge join prior to hashjoin
-      // for now let's support inner join first, then add outer join
       case ExtractEquiJoinKeys(Inner, leftKeys, rightKeys, condition, left, right)
         if sqlContext.conf.sortMergeJoinEnabled && RowOrdering.isOrderable(leftKeys) =>
         val mergeJoin =
@@ -115,6 +120,8 @@ private[sql] abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
           leftKeys, rightKeys, buildSide, planLater(left), planLater(right))
         condition.map(Filter(_, hashJoin)).getOrElse(hashJoin) :: Nil
 
+      // --- Outer joins --------------------------------------------------------------------------
+
       case ExtractEquiJoinKeys(
              LeftOuter, leftKeys, rightKeys, condition, left, CanBroadcast(right)) =>
         joins.BroadcastHashOuterJoin(
@@ -125,10 +132,22 @@ private[sql] abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
         joins.BroadcastHashOuterJoin(
           leftKeys, rightKeys, RightOuter, condition, planLater(left), planLater(right)) :: Nil
 
+      case ExtractEquiJoinKeys(LeftOuter, leftKeys, rightKeys, condition, left, right)
+        if sqlContext.conf.sortMergeJoinEnabled && RowOrdering.isOrderable(leftKeys) =>
+        joins.SortMergeOuterJoin(
+          leftKeys, rightKeys, LeftOuter, condition, planLater(left), planLater(right)) :: Nil
+
+      case ExtractEquiJoinKeys(RightOuter, leftKeys, rightKeys, condition, left, right)
+        if sqlContext.conf.sortMergeJoinEnabled && RowOrdering.isOrderable(leftKeys) =>
+        joins.SortMergeOuterJoin(
+          leftKeys, rightKeys, RightOuter, condition, planLater(left), planLater(right)) :: Nil
+
       case ExtractEquiJoinKeys(joinType, leftKeys, rightKeys, condition, left, right) =>
         joins.ShuffledHashOuterJoin(
           leftKeys, rightKeys, joinType, condition, planLater(left), planLater(right)) :: Nil
 
+      // --- Cases where this strategy does not apply ---------------------------------------------
+
       case _ => Nil
     }
   }

sql/core/src/main/scala/org/apache/spark/sql/execution/joins/BroadcastNestedLoopJoin.scala

@@ -65,8 +65,9 @@ case class BroadcastNestedLoopJoin(
         left.output.map(_.withNullability(true)) ++ right.output
       case FullOuter =>
         left.output.map(_.withNullability(true)) ++ right.output.map(_.withNullability(true))
-      case _ =>
-        left.output ++ right.output
+      case x =>
+        throw new IllegalArgumentException(
+          s"BroadcastNestedLoopJoin should not take $x as the JoinType")
     }
   }