[SPARK-31478][CORE]Call StopExecutor before killing executors
#28254
Conversation
|
cc @dongjoon-hyun @vanzin |
|
Since this PR aims reverts the original patch of SPARK-29152 and add a new way, I believe @vanzin 's option is important to this. |
|
ok to test |
|
Test build #121459 has finished for PR 28254 at commit
|
| val killExecutors: Boolean => Future[Boolean] = | ||
| if (executorsToKill.nonEmpty) { | ||
| executorsToKill.foreach(id => | ||
| executorDataMap.get(id).foreach(_.executorEndpoint.send(StopExecutor))) |
There was a problem hiding this comment.
Can we guarantee that stop is called before kill in this way?
There was a problem hiding this comment.
StopExecutor may arrive at executor after kill arrive at worker/container due to network delay, isn't it possible?
|
Sorry, don't really have the cycles for a detailed review. But if you remove the shutdown hook, you'll be missing certain executor exit scenarios, like a cluster manager killing the executor. |
|
For the time being we are continuing with the current approach only. Closing this PR. |
5 participants
What changes were proposed in this pull request?
StopExecutorcall to executors before killing them.A similar patch is tested here #26901
Why are the changes needed?
When executors do not goes down gracefully, their
stop()method is never called. To solve this problem, a shutdown hook was used to execute thestop()method.Instead of forcing a shutdown hook, we should just add a
StopExecutorcall to executors before they are killed.Does this PR introduce any user-facing change?
No.
How was this patch tested?
Manually