[ 
https://issues.apache.org/jira/browse/YARN-9007?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16682851#comment-16682851
 ] 

Tao Yang edited comment on YARN-9007 at 11/11/18 11:53 AM:
-----------------------------------------------------------

Add "for queue and reserved container preemption" to the end of this issue 
name. This issue is proposed to skip selecting OPPORTUNISTIC containers for the 
policies which want to balance GUARANTEED resource  in preemption monitor.
{quote}
Few high level thoughts.
1. IMO, ExecutionType.OPPORTUNISTIC containers could be the first candidates 
for any preemption. Rather than selecting GUARANTEED containers before 
consuming OPPORTUNISTIC seems not best choice as per me.
2. On the flip side, i know OPPORTUNISTIC containers has a different lifecycle 
when more demand comes on the node. But to me, both preemption and existing 
model of upgrading/downgrading ExecutionType need to coexist together. So could 
we discuss more on that line.
in either of above case, we need more discussion.
{quote}
[~sunilg], I agree with your thoughts. Sorry about not clarify this issue 
before.


was (Author: tao yang):
Add "for queue and reserved container preemption" to the end of this issue 
name. This issue is proposed to skip selecting OPPORTUNISTIC containers for the 
policies which want to balance GUARANTEED resource  in preemption monitor.

> CS preemption monitor should only select GUARANTEED containers as candidates 
> for queue and reserved container preemption
> ------------------------------------------------------------------------------------------------------------------------
>
>                 Key: YARN-9007
>                 URL: https://issues.apache.org/jira/browse/YARN-9007
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: capacityscheduler
>    Affects Versions: 3.2.1
>            Reporter: Tao Yang
>            Assignee: Tao Yang
>            Priority: Major
>         Attachments: YARN-9007.001.patch
>
>
> Currently CS preemption monitor doesn't consider execution type of 
> containers, so OPPORTUNISTIC containers maybe selected and killed without 
> effect.
> In some scenario with OPPORTUNISTIC containers, not even preemption can't 
> work properly to balance resources, but also some apps with OPPORTUNISTIC 
> containers maybe effected and unable to work.



--
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

Reply via email to