Github user jinxing64 commented on the issue: https://github.com/apache/spark/pull/16620 @kayousterhout @squito @markhamstra Thanks a lot for for the comments. I've already refined accordingly. I still have one concern: > If this is a correct description, Iâd argue that (5) is the problem: that when ShuffleMapTask2 finishes, we should not be updating a bunch of state in the DAGScheduler saying that thereâs output ready as a result. If Iâm understanding correctly, thereâs a relatively simple fix to this problem: In DAGScheduler.scala, in handleTaskCompletion, we should exit (and not update any state) when the task is from an earlier stage attempt thatâs not the current active attempt. This can be done by changing the if-statement on line 1141 to include: || stageIdToStage(task.stageId).latestInfo.attemptId != task.stageAttemptId With above, are we ignoring all the results from old stage attempts? As @squito mentioned: > It also can potentially improve performance, since you may submit downstream stages more quickly, rather than waiting for all tasks in the active taskset to complete. Is it maybe beneficial to add up the result from old stage attempts.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org