mridulm commented on a change in pull request #30164:
URL: https://github.com/apache/spark/pull/30164#discussion_r519333540
##########
File path:
resource-managers/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnSchedulerBackend.scala
##########
@@ -161,6 +170,27 @@ private[spark] abstract class YarnSchedulerBackend(
totalRegisteredExecutors.get() >= totalExpectedExecutors *
minRegisteredRatio
}
+ override def getMergerLocations(
+ numPartitions: Int,
+ resourceProfileId: Int): Seq[BlockManagerId] = {
+ // Currently this is naive way of calculating numMergersNeeded for a
stage. In future,
+ // we can use better heuristics to calculate numMergersNeeded for a stage.
+ val tasksPerExecutor = sc.resourceProfileManager
+ .resourceProfileFromId(resourceProfileId).maxTasksPerExecutor(sc.conf)
+ val numMergersNeeded = math.min(
+ math.max(1, math.ceil(numPartitions / tasksPerExecutor).toInt),
maxNumExecutors)
+ val minMergersThreshold = math.max(minMergersStaticThreshold,
+ math.floor(numMergersNeeded * minMergersThresholdRatio).toInt)
Review comment:
The name `minMergersThreshold` is a bit confusing - can you rename it ?
With `minMergersThreshold `, I would expect the minimum required to be
enforced irrespective of other conditions. If mergers < threshold, no push
based shuffle. This could be either due to lack of available mergers, due to a
very small stage or other reasons.
I am guessing the intention is not to do this, which is why more clarity in
variable name should help. Thoughts ?
----------------------------------------------------------------
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
