[
https://issues.apache.org/jira/browse/YARN-1292?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13792063#comment-13792063
]
Bikas Saha commented on YARN-1292:
----------------------------------
This can be achieved in a backwards compatible manner in the following way
1) StartContainer request will have a new flag that says whether the container
is attached to a process or not. Default value is true for back-compat.
2) If the above flag is false then the container is completed on the NM only
when
a) the RM terminates the container (this currently happens today)
b) when the AM call StopContainer on that (this is currently supported)
The main change in the NM would be to not trigger end of container, ie keep the
container in a running state, when there is no process associated with the
container.
3) Create a new api called startProcess() that can be used to launch a new
process in a container. NM can dis-allow starting a process while a process is
already running for the first cut. This API would be secured using existing
AMNM token.
No changes are expected to be needed in the RM since the NM will continue to
report this container as running to the RM. This should be a fairly localised
NM-only change.
> De-link container life cycle from the process it runs
> -----------------------------------------------------
>
> Key: YARN-1292
> URL: https://issues.apache.org/jira/browse/YARN-1292
> Project: Hadoop YARN
> Issue Type: Improvement
> Affects Versions: 2.1.1-beta
> Reporter: Bikas Saha
>
> Currently, a container is considered done when its OS process exits. This
> makes it cumbersome for apps to be able to reuse containers for different
> processes. Long running daemons may want to run in the same containers as the
> previous versions. So eg. is an hbase region server crashes/upgraded it would
> want to restart in the same container where everything it needs would already
> be warm and ready.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
