Github user arunmahadevan commented on a diff in the pull request:
https://github.com/apache/spark/pull/21385#discussion_r190320670
--- Diff:
sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/continuous/shuffle/UnsafeRowReceiver.scala
---
@@ -56,20 +69,71 @@ private[shuffle] class UnsafeRowReceiver(
override def receiveAndReply(context: RpcCallContext):
PartialFunction[Any, Unit] = {
case r: UnsafeRowReceiverMessage =>
- queue.put(r)
+ queues(r.writerId).put(r)
context.reply(())
}
override def read(): Iterator[UnsafeRow] = {
new NextIterator[UnsafeRow] {
- override def getNext(): UnsafeRow = queue.take() match {
- case ReceiverRow(r) => r
- case ReceiverEpochMarker() =>
- finished = true
- null
+ // An array of flags for whether each writer ID has gotten an epoch
marker.
+ private val writerEpochMarkersReceived =
Array.fill(numShuffleWriters)(false)
+
+ private val executor =
Executors.newFixedThreadPool(numShuffleWriters)
--- End diff --
Are we creating threads equal to num writers here and do we need it? I
assume there would be already the RDD task thread per partition that's invoking
`getNext` and cant we return the row from the right queue(s) there ? My concern
is that these threads are going to bloat up the total threads to m*n (rdd
partitions * shuffle writers) and this may not be very scalable considering the
amount of context switch.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
