[ https://issues.apache.org/jira/browse/YARN-671?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13802878#comment-13802878 ]
Jason Lowe commented on YARN-671: --------------------------------- Note that simply waiting for containers to complete doesn't address auxiliary services like MapReduce's ShuffleHandler that want to serve up data after containers have completed. If we simply wait until containers have completed and stop the NM immediately afterwards then we can cause fetch failures leading to a relaunch of map tasks that ran on that node. In that case we would have been better off killing the map task immediately, as the application would have recovered faster. Therefore without more intimate knowledge of what an auxiliary service is really doing with respect to the containers that ran on a node, we may have to generalize the timeout to all applications have completed that have run at least one container on the node. > Add an interface on the RM to move NMs into a maintenance state > --------------------------------------------------------------- > > Key: YARN-671 > URL: https://issues.apache.org/jira/browse/YARN-671 > Project: Hadoop YARN > Issue Type: Sub-task > Affects Versions: 2.0.4-alpha > Reporter: Siddharth Seth > Assignee: Siddharth Seth > -- This message was sent by Atlassian JIRA (v6.1#6144)