vanzin commented on a change in pull request #24817: [SPARK-27963][core] Allow
dynamic allocation without a shuffle service.
URL: https://github.com/apache/spark/pull/24817#discussion_r298375961
##########
File path:
core/src/main/scala/org/apache/spark/scheduler/dynalloc/ExecutorMonitor.scala
##########
@@ -124,6 +149,98 @@ private[spark] class ExecutorMonitor(
def pendingRemovalCount: Int = executors.asScala.count { case (_, exec) =>
exec.pendingRemoval }
+ override def onJobStart(event: SparkListenerJobStart): Unit = {
+ if (!shuffleTrackingEnabled) {
+ return
+ }
+
+ val shuffleStages = event.stageInfos.flatMap { s =>
+ s.shuffleDepId.toSeq.map { shuffleId =>
+ s.stageId -> shuffleId
+ }
+ }
+
+ var updateExecutors = false
+ shuffleStages.foreach { case (stageId, shuffle) =>
+ val jobIDs = shuffleToActiveJobs.get(shuffle) match {
+ case Some(jobs) =>
+ // If a shuffle is being re-used, we need to re-scan the executors
and update their
+ // tracker with the information that the shuffle data they're
storing is in use.
+ logDebug(s"Reusing shuffle $shuffle in job ${event.jobId}.")
+ updateExecutors = true
+ jobs
+
+ case _ =>
+ logDebug(s"Registered new shuffle $shuffle (from stage $stageId).")
+ val jobs = new mutable.ArrayBuffer[Int]()
+ shuffleToActiveJobs(shuffle) = jobs
+ jobs
+ }
+ jobIDs += event.jobId
+ }
+
+ if (updateExecutors) {
+ val activeShuffleIds = shuffleStages.map(_._2).toSeq
Review comment:
While what you say is true in isolation, the check for `if
(!exec.hasActiveShuffle) {` below makes sure the end result is correct. When
the second jobs starts, executor X has an active shuffle, so this code would
not change that state.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
