Details
-
Improvement
-
Status: Resolved
-
Minor
-
Resolution: Incomplete
-
2.2.0
-
None
Description
Yarn provides max attempt configuration for applications running on it, application has the chance to retry itself when failed. In the current Spark code, no matter which failure AM occurred and if the failure doesn't reach to the max attempt, RM will restart AM, this is not reasonable for some cases if this issue is coming from AM itself, like user code failure, OOM, Spark issue, executor failures, in large chance the reattempt of AM will meet this issue again. Only when AM is failed due to external issue like crash, process kill, NM failure, then AM should retry again.
So here propose to improve this reattempt mechanism to only retry when it meets external issues.