[
https://issues.apache.org/jira/browse/YARN-4725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15184821#comment-15184821
]
Junping Du commented on YARN-4725:
----------------------------------
For the same requirement, can we implement this in another way - adding
stickiness to first attempt NM for service container in RM scheduling, rather
than NM launch? Also, in another opinion, service container could need higher
bar for quality of NM that may need to blacklist NMs that cause previous
failure in follow-up running.
> [Umbrella] Auto-restart of containers
> --------------------------------------
>
> Key: YARN-4725
> URL: https://issues.apache.org/jira/browse/YARN-4725
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Vinod Kumar Vavilapalli
>
> See overview doc at YARN-4692, copying the sub-section to track all related
> efforts.
> Today, when a container (process-tree) dies, NodeManager assumes that the
> container’s allocation is also expired, and reports accordingly to the
> ResourceManager which then releases the allocation. For service containers,
> this is undesirable in many cases. Long running containers may exit for
> various reasons, crash and need to restart but forcing them to go through the
> complete scheduling cycle, resource localization etc is both unnecessary and
> expensive. (Task) For services it will be good to have NodeManagers
> automatically restart containers. This looks a lot like inittab /
> daemontools at the system level.
> We will need to enable app-specific policies (very similar to the handling
> of AM restarts at YARN level) for restarting containers automatically but
> limit such restarts if a container dies too often in a short interval of time.
> YARN-3998 is an existing ticket that looks at some if not all of this
> functionality.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
