[
https://issues.apache.org/jira/browse/YARN-2489?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16772289#comment-16772289
]
Íñigo Goiri commented on YARN-2489:
-----------------------------------
I added a patch to YARN-999 which pretty much covers the description of this
JIRA and the actual killing when we over commit.
What we do there is to change the resources of the NM and then after the over
commit time out happens, we kill.
Another option would be to have in this JIRA the mechanism to wait X seconds to
change the resources and YARN-999 to just kill when we go negative.
I think the current approach in YARN-999 covers the functionality better as it
would allow to reduce the size of the NM and wait forever until containers are
drained while showing the change in resources.
> ResouceOption's overcommitTimeout should be respected during resource update
> on NM
> ----------------------------------------------------------------------------------
>
> Key: YARN-2489
> URL: https://issues.apache.org/jira/browse/YARN-2489
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: graceful, nodemanager, scheduler
> Reporter: Junping Du
> Priority: Major
>
> The ResourceOption to update NM's resource has two properties: Resource and
> OvercommitTimeout. The later property is used to guarantee resource is
> withdrawn after timeout is hit if resource is reduced to a value and current
> resource consumption exceeds the new value. It currently use default value -1
> which means no timeout, and we should make this property work when updating
> NM resource.
--
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
