[ https://issues.apache.org/jira/browse/MAPREDUCE-5888?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Jason Lowe updated MAPREDUCE-5888: ---------------------------------- Attachment: MAPREDUCE-5888.patch Quick patch to fix the issue. Manually tested it with a fail job and saw that the MRAppMaster hung after unregistering without the change and does not hang with the patch. > Failed job leaves hung AM after it unregisters > ----------------------------------------------- > > Key: MAPREDUCE-5888 > URL: https://issues.apache.org/jira/browse/MAPREDUCE-5888 > Project: Hadoop Map/Reduce > Issue Type: Bug > Components: mr-am > Affects Versions: 2.2.0 > Reporter: Jason Lowe > Assignee: Jason Lowe > Attachments: MAPREDUCE-5888.patch > > > When a job fails the AM hangs during shutdown. A non-daemon thread pool > executor thread prevents the JVM teardown from completing, and the AM lingers > on the cluster for the AM expiry interval in the FINISHING state until > eventually the RM expires it and kills the container. If application limits > on the queue are relatively low (e.g.: small queue or small cluster) this can > cause unnecessary delays in resource scheduling on the cluster. -- This message was sent by Atlassian JIRA (v6.2#6252)