Github user nongli commented on a diff in the pull request:
https://github.com/apache/spark/pull/11328#discussion_r54163596
--- Diff:
sql/core/src/main/scala/org/apache/spark/sql/execution/joins/BroadcastNestedLoopJoin.scala
---
@@ -51,125 +51,254 @@ case class BroadcastNestedLoopJoin(
}
private[this] def genResultProjection: InternalRow => InternalRow = {
- UnsafeProjection.create(schema)
+ if (joinType == LeftSemi) {
+ UnsafeProjection.create(output, output)
+ } else {
+ // Always put the stream side on left to simplify implementation
+ UnsafeProjection.create(output, streamed.output ++ broadcast.output)
+ }
}
override def outputPartitioning: Partitioning =
streamed.outputPartitioning
override def output: Seq[Attribute] = {
joinType match {
+ case Inner =>
+ left.output ++ right.output
case LeftOuter =>
left.output ++ right.output.map(_.withNullability(true))
case RightOuter =>
left.output.map(_.withNullability(true)) ++ right.output
case FullOuter =>
left.output.map(_.withNullability(true)) ++
right.output.map(_.withNullability(true))
- case Inner =>
- // TODO we can avoid breaking the lineage, since we union an empty
RDD for Inner Join case
- left.output ++ right.output
- case x => // TODO support the Left Semi Join
+ case LeftSemi =>
+ left.output
+ case x =>
throw new IllegalArgumentException(
s"BroadcastNestedLoopJoin should not take $x as the JoinType")
}
}
- @transient private lazy val boundCondition =
- newPredicate(condition.getOrElse(Literal(true)), left.output ++
right.output)
+ @transient private lazy val boundCondition = {
+ if (condition.isDefined) {
+ newPredicate(condition.get, streamed.output ++ broadcast.output)
+ } else {
+ (r: InternalRow) => true
+ }
+ }
- protected override def doExecute(): RDD[InternalRow] = {
- val numOutputRows = longMetric("numOutputRows")
+ private def innerJoin(relation: Broadcast[Array[InternalRow]]):
RDD[InternalRow] = {
+ streamed.execute().mapPartitionsInternal { streamedIter =>
+ val buildRows = relation.value
+ val joinedRow = new JoinedRow
- val broadcastedRelation =
broadcast.executeBroadcast[Array[InternalRow]]()
+ streamedIter.flatMap { streamedRow =>
+ val joinedRows = buildRows.iterator.map(r =>
joinedRow(streamedRow, r))
+ if (condition.isDefined) {
+ joinedRows.filter(boundCondition)
+ } else {
+ joinedRows
+ }
+ }
+ }
+ }
- /** All rows that either match both-way, or rows from streamed joined
with nulls. */
- val matchesOrStreamedRowsWithNulls = streamed.execute().mapPartitions
{ streamedIter =>
- val relation = broadcastedRelation.value
+ private def outerJoin(relation: Broadcast[Array[InternalRow]]):
RDD[InternalRow] = {
+ streamed.execute().mapPartitionsInternal { streamedIter =>
+ val buildRows = relation.value
+ val joinedRow = new JoinedRow
+ val nulls = new GenericMutableRow(broadcast.output.size)
+
+ // Returns an iterator to avoid copy the rows.
+ new Iterator[InternalRow] {
+ // current row from stream side
+ private var streamRow: InternalRow = null
+ // have found a match for current row or not
+ private var foundMatch: Boolean = false
+ // the matched result row
+ private var resultRow: InternalRow = null
+ // the next index of buildRows to try
+ private var nextIndex: Int = 0
- val matchedRows = new CompactBuffer[InternalRow]
- val includedBroadcastTuples = new BitSet(relation.length)
+ private def findNextMatch(): Boolean = {
+ if (streamRow == null) {
+ if (!streamedIter.hasNext) {
+ return false
+ }
+ streamRow = streamedIter.next()
+ nextIndex = 0
+ foundMatch = false
+ }
+ while (nextIndex < buildRows.length) {
+ resultRow = joinedRow(streamRow, buildRows(nextIndex))
+ nextIndex += 1
+ if (boundCondition(resultRow)) {
+ foundMatch = true
+ return true
+ }
+ }
+ if (!foundMatch) {
+ resultRow = joinedRow(streamRow, nulls)
+ streamRow = null
+ true
+ } else {
+ resultRow = null
+ streamRow = null
+ findNextMatch()
+ }
+ }
+
+ override def hasNext(): Boolean = {
+ resultRow != null || findNextMatch()
+ }
+ override def next(): InternalRow = {
+ val r = resultRow
+ resultRow = null
+ r
+ }
+ }
+ }
+ }
+
+ private def leftSemiJoin(relation: Broadcast[Array[InternalRow]]):
RDD[InternalRow] = {
+ assert(buildSide == BuildRight)
+ streamed.execute().mapPartitionsInternal { streamedIter =>
+ val buildRows = relation.value
val joinedRow = new JoinedRow
- val leftNulls = new GenericMutableRow(left.output.size)
- val rightNulls = new GenericMutableRow(right.output.size)
- val resultProj = genResultProjection
+ if (condition.isDefined) {
+ streamedIter.filter(l =>
+ buildRows.exists(r => boundCondition(joinedRow(l, r)))
+ )
+ } else {
+ streamedIter.filter(r => !buildRows.isEmpty)
+ }
+ }
+ }
+
+ /**
+ * The implementation these joins:
--- End diff --
"for these joins"
---
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]
