[ https://issues.apache.org/jira/browse/YARN-4680?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15138248#comment-15138248 ]
Xuan Gong commented on YARN-4680: --------------------------------- The proposal is to fix the memory leak inside the ATS v1.5 write. It includes * Lazy start the timerTasks. We do not start flushTimer and cleanInActiveFDsTimer until we actually have the ats v1.5 write operation. * Add a configuration: yarn.timeline-service.client.timer-task-retain-secs, and schedule another task: monitorTimerTask in cleanInActiveFDsTimer which will check System.getCurrentTimeStamp - timeStampOfLastWrite >= this.timerTaskRetainTTL(yarn.timeline-service.client.timer-task-retain-secs). If it is true which mean in a configured period of time, we do not have any write operation, we would cancel all the timers. > TimerTasks leak in ATS V1.5 Writer > ---------------------------------- > > Key: YARN-4680 > URL: https://issues.apache.org/jira/browse/YARN-4680 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelineserver > Reporter: Xuan Gong > Assignee: Xuan Gong > > We have seen TimerTasks leak which could cause application server done (such > as oozie server done due to too many active threads) > Although we have fixed some potentially leak situations in upper application > level, such as > https://issues.apache.org/jira/browse/MAPREDUCE-6618 > https://issues.apache.org/jira/browse/MAPREDUCE-6621, we still can not > guarantee that we fixed the issue. -- This message was sent by Atlassian JIRA (v6.3.4#6332)