[
https://issues.apache.org/jira/browse/YARN-906?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13711962#comment-13711962
]
Zhijie Shen commented on YARN-906:
----------------------------------
Thanks Omkar for the feedback.
bq. an alternative solution which seems most logical to me is that what if we
send the same event from the place where we are canceling thread and expect
/ignore additional event at KILLING state...
Thought about it before. The problem is that we don't need to send
CONTAINER_KILLED_ON_REQUEST every time when canceling call(). Full logic needs
to be implemented here to decide whether CONTAINER_KILLED_ON_REQUEST needs to
be emitted. So why not letting call() move on to achieve the same thing?
bq. what surprises me here is our single dispatcher thread model..
IMHO, except non-optimized performance, single dispatcher is already friendly
for event driven model. Multiple event queues / dispatchers will make tracing
the total order of events exponentially more complex. For example, the problem
in this jira is difficult to trace or reproduce because
CONTAINER_KILLED_ON_REQUEST is emitted in another thread instead of the
dispatcher thread. If call() is not scheduled in another thread, but runs on
the dispatcher thread, the problem will not happen.
> TestNMClient.testNMClientNoCleanupOnStop fails occasionally
> -----------------------------------------------------------
>
> Key: YARN-906
> URL: https://issues.apache.org/jira/browse/YARN-906
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Zhijie Shen
> Assignee: Zhijie Shen
> Attachments: YARN-906.1.patch
>
>
> See
> https://builds.apache.org/job/PreCommit-YARN-Build/1435//testReport/org.apache.hadoop.yarn.client.api.impl/TestNMClient/testNMClientNoCleanupOnStop/
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
