[
https://issues.apache.org/jira/browse/YARN-433?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14632320#comment-14632320
]
zhihai xu commented on YARN-433:
--------------------------------
Thanks for the patch [~xgong]! The patch looks good to me except two nits:
# It may be better to use {{containerId}} instead of
{{remoteContainer.getContainerId()}} as the parameter of
{{containerAllocationExpirer.unregister}} since they are equivalent.
# Since we do {{containerAllocationExpirer.unregister}} for the completed
containers in {{RMNodeImpl#handleContainerStatus}} earlier, Maybe we can remove
the redundant code {{containerAllocationExpirer.unregister}} at
{{RMContainerImpl#ContainerFinishedAtAcquiredState}}. After we remove this code
at {{RMContainerImpl#ContainerFinishedAtAcquiredState}},
ContainerFinishedAtAcquiredState will be the same as FinishedTransition, So we
can replace ContainerFinishedAtAcquiredState with FinishedTransition and remove
ContainerFinishedAtAcquiredState.
{code}
.addTransition(RMContainerState.ACQUIRED, RMContainerState.COMPLETED,
RMContainerEventType.FINISHED, new FinishedTransition())
{code}
> When RM is catching up with node updates then it should not expire acquired
> containers
> --------------------------------------------------------------------------------------
>
> Key: YARN-433
> URL: https://issues.apache.org/jira/browse/YARN-433
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: resourcemanager
> Reporter: Bikas Saha
> Assignee: Xuan Gong
> Attachments: YARN-433.1.patch, YARN-433.2.patch, YARN-433.3.patch
>
>
> RM expires containers that are not launched within some time of being
> allocated. The default is 10mins. When an RM is not keeping up with node
> updates then it may not be aware of new launched containers. If the expire
> thread fires for such containers then the RM can expire them even though they
> may have launched.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
