[ https://issues.apache.org/jira/browse/SPARK-22876?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16991752#comment-16991752 ]
Andrew White commented on SPARK-22876: -------------------------------------- This issue is most certainly unresolved, and the documentation is misleading at best. [~choojoyq] - perhaps you could share your code on how you deal with this in general? The general solution seems non-obvious to me given all of the intricacies of Spark + Yarn interactions such as shutdown hooks, handling signals, client vs cluster mode. Maybe I'm over thinking the solution. This feature is critical for supporting long-running applications. > spark.yarn.am.attemptFailuresValidityInterval does not work correctly > --------------------------------------------------------------------- > > Key: SPARK-22876 > URL: https://issues.apache.org/jira/browse/SPARK-22876 > Project: Spark > Issue Type: Bug > Components: YARN > Affects Versions: 2.2.0 > Environment: hadoop version 2.7.3 > Reporter: Jinhan Zhong > Priority: Minor > Labels: bulk-closed > > I assume we can use spark.yarn.maxAppAttempts together with > spark.yarn.am.attemptFailuresValidityInterval to make a long running > application avoid stopping after acceptable number of failures. > But after testing, I found that the application always stops after failing n > times ( n is minimum value of spark.yarn.maxAppAttempts and > yarn.resourcemanager.am.max-attempts from client yarn-site.xml) > for example, following setup will allow the application master to fail 20 > times. > * spark.yarn.am.attemptFailuresValidityInterval=1s > * spark.yarn.maxAppAttempts=20 > * yarn client: yarn.resourcemanager.am.max-attempts=20 > * yarn resource manager: yarn.resourcemanager.am.max-attempts=3 > And after checking the source code, I found in source file > ApplicationMaster.scala > https://github.com/apache/spark/blob/master/resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala#L293 > there's a ShutdownHook that checks the attempt id against the maxAppAttempts, > if attempt id >= maxAppAttempts, it will try to unregister the application > and the application will finish. > is this a expected design or a bug? -- This message was sent by Atlassian Jira (v8.3.4#803005) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org