Ngone51 commented on a change in pull request #30650:
URL: https://github.com/apache/spark/pull/30650#discussion_r553094418
##########
File path:
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
##########
@@ -661,35 +667,52 @@ private[spark] class TaskSchedulerImpl(
}
if (launchedAnyTask && taskSet.isBarrier) {
+ val barrierPendingLaunchTasks =
taskSet.barrierPendingLaunchTasks.values.toArray
// Check whether the barrier tasks are partially launched.
- // TODO SPARK-24818 handle the assert failure case (that can happen
when some locality
- // requirements are not fulfilled, and we should revert the launched
tasks).
- if (addressesWithDescs.size != taskSet.numTasks) {
- val errorMsg =
- s"Fail resource offers for barrier stage ${taskSet.stageId}
because only " +
- s"${addressesWithDescs.size} out of a total number of
${taskSet.numTasks}" +
- s" tasks got resource offers. This happens because barrier
execution currently " +
- s"does not work gracefully with delay scheduling. We highly
recommend you to " +
- s"disable delay scheduling by setting spark.locality.wait=0 as
a workaround if " +
- s"you see this error frequently."
- logWarning(errorMsg)
- taskSet.abort(errorMsg)
- throw new SparkException(errorMsg)
- }
+ if (barrierPendingLaunchTasks.length != taskSet.numTasks) {
+ barrierPendingLaunchTasks.foreach { task =>
+ // revert all assigned resources
+ availableCpus(task.assignedOfferIndex) += task.assignedCores
+ task.assignedResources.foreach { case (rName, rInfo) =>
+
availableResources(task.assignedOfferIndex)(rName).appendAll(rInfo.addresses)
+ }
+ // re-add the task to the schedule pending list
+ taskSet.addPendingTask(task.index)
Review comment:
thanks for pointing this out. I'd like to take a look at rack resolving
and see if we could have a way to avoid redundant resolving.
----------------------------------------------------------------
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]
