[ https://issues.apache.org/jira/browse/YARN-8789?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16709605#comment-16709605 ]
Wilfred Spiegelenburg commented on YARN-8789: --------------------------------------------- First: I would expect that the change would be fully tested and thus the behaviour with a limited queue would be known and described. Task failures are probably more acceptable. Are we really still seeing them with the change from MAPREDUCE-5124 applied? If not then making this change is not really warranted. Before we go further and make a change like this I would also test the behaviour. What happens when the queue is full. Looking at the patch there is far more change than needed: the current queue can be limited and just that change would be far less impactful. The logic for taking an event is also changed which I don't think is needed either. Going back to just the basic change of limiting the queue after we find that it is needed would be a better approach. Based on that quick analysis I would say this is not an acceptable change in its current form. > Add BoundedQueue to AsyncDispatcher > ----------------------------------- > > Key: YARN-8789 > URL: https://issues.apache.org/jira/browse/YARN-8789 > Project: Hadoop YARN > Issue Type: Improvement > Components: applications > Affects Versions: 3.2.0 > Reporter: BELUGA BEHR > Assignee: BELUGA BEHR > Priority: Major > Attachments: YARN-8789.1.patch, YARN-8789.10.patch, > YARN-8789.12.patch, YARN-8789.14.patch, YARN-8789.2.patch, YARN-8789.3.patch, > YARN-8789.4.patch, YARN-8789.5.patch, YARN-8789.6.patch, YARN-8789.7.patch, > YARN-8789.7.patch, YARN-8789.8.patch, YARN-8789.9.patch > > > I recently came across a scenario where an MR ApplicationMaster was failing > with an OOM exception. It had many thousands of Mappers and thousands of > Reducers. It was noted that in the logging that the event-queue of > {{AsyncDispatcher}} had a very large number of item in it and was seemingly > never decreasing. > I started looking at the code and thought it could use some clean up, > simplification, and the ability to specify a bounded queue so that any > incoming events are throttled until they can be processed. This will protect > the ApplicationMaster from a flood of events. > Logging Message: > Size of event-queue is xxx -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org