Ngone51 commented on a change in pull request #22806: [SPARK-25250][CORE] :
Late zombie task completions handled correctly even before new taskset launched
URL: https://github.com/apache/spark/pull/22806#discussion_r247012072
##########
File path:
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
##########
@@ -286,6 +286,44 @@ private[spark] class TaskSchedulerImpl(
}
}
+ /**
+ * SPARK-25250: Whenever any Task gets successfully completed, we simply
mark the
+ * corresponding partition id as completed in all attempts for that
particular stage and
+ * additionally, for a Result Stage, we also kill the remaining task
attempts running on the
+ * same partition. As a result, we do not see any Killed tasks due to
+ * TaskCommitDenied Exceptions showing up in the UI. When this method is
called from
+ * DAGScheduler.scala on a task completion event being fired, it is assumed
that the new
+ * TaskSet has already been created and registered. However, a small
possibility does exist
+ * that when this method gets called, possibly the new TaskSet might have
not been added
+ * to taskSetsByStageIdAndAttempt. In such a case, we might still hit the
same issue. However,
+ * the above scenario has not yet been reproduced.
+ */
+ override def completeTasks(partitionId: Int, stageId: Int, killTasks:
Boolean): Unit = {
+ taskSetsByStageIdAndAttempt.getOrElse(stageId, Map()).values.foreach { tsm
=>
+ tsm.partitionToIndex.get(partitionId) match {
+ case Some(index) =>
+ tsm.markPartitionAsAlreadyCompleted(index)
Review comment:
Hmmm... I'm not saying we can't mark `successful(index) = true`, but we
should not only mark `successful(index) = true` but also mark `isZombie=true`
for some cases. Because, a `tsm` can not be finished if it is not zombie.
And to be clarify, again, the key point we're arguing is none business of
whether to kill the tasks. We'll kill the tasks after calling
`markPartitionAsAlreadyCompleted()`, I promise.
> If `tsm` is active and the task is pending to run, I think killing the
task is a no-op, but maybe better to not send the KillTask request to avoid
overhead. `successful(index) = true` is needed here to finish the `tsm`.
We should not only mark `successful(index) = true`, but also increase
`tasksSuccessful`(this may has a little affect on blacklist) and even `mark the
`tsm` as zombie if the task is the last successful task for the `tsm`.
otherwise, it can not be finished .
>if `tsm` is active and the task is running. Killing the task will probably
save resource, and `successful(index) = true` is needed here to finish the
`tsm`.
Same as above.
> if `tsm` is zoombie, `successful(index) = true` is a no-op(but no
overhead) as a zombie `tsm` will never be marked as finished. Killing the task
may have benefit if the task is still running.
As I mentioned previously, mark `successful(index) = true` is unnecessary.
So, I'm fine whether we do this or not.
Correct me if I'm wrong, thanks, @cloud-fan .
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on 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]
