[ https://issues.apache.org/jira/browse/HIVE-14608?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15472433#comment-15472433 ]
Sergey Shelukhin commented on HIVE-14608: ----------------------------------------- It mostly applies to reducers actually. The main problem is as indicated in the description - for whatever reason we don't remove nodes from the node list when they die. The per-node assignment explicitly checks active set to get around that(?) but the other path doesn't... so reducers with no location preference will be sent to dead nodes, potentially > LLAP: slow scheduling due to LlapTaskScheduler not removing nodes on kill > -------------------------------------------------------------------------- > > Key: HIVE-14608 > URL: https://issues.apache.org/jira/browse/HIVE-14608 > Project: Hive > Issue Type: Bug > Reporter: Sergey Shelukhin > Assignee: Sergey Shelukhin > Priority: Critical > Attachments: HIVE-14608.patch > > > See comments; this can result in a slowdown esp. if some critical task gets > unlucky. > {noformat} > public void workerNodeRemoved(ServiceInstance serviceInstance) { > // FIXME: disabling this for now > // instanceToNodeMap.remove(serviceInstance.getWorkerIdentity()); > {noformat} -- This message was sent by Atlassian JIRA (v6.3.4#6332)