[
https://issues.apache.org/jira/browse/YUNIKORN-402?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17196889#comment-17196889
]
Kinga Marton commented on YUNIKORN-402:
---------------------------------------
[~wwei], unfortunately with the actual architecture is not so easy to add that
check in case of allocation removal.
The actual workflow is the following one:
- the cache is sending a {{SchedulerAllocationUpdatesEvent}} to the scheduler:
[https://github.com/apache/incubator-yunikorn-core/blob/master/pkg/cache/cluster_info.go#L310]
- the scheduler first will process the event then will send a
{{ReleaseAllocationsEvent}} event to the cache to remove the allocation from
the application_info:
[https://github.com/apache/incubator-yunikorn-core/blob/master/pkg/scheduler/scheduler.go#L400-L402].
At this point in the scheduler, the scheduling application still has the
allocation. This will be removed only in the cache side. So here we don't know
if we will still have some allocated resources after removing the ones from the
request. The processing in the cache side is async, so if we would change the
order would not help in this case. When removing the allocation from the
application_info in the cache side, there we don't know if we have some pending
asks.
I think that after merging the cache and scheduler side, solving this issue
would be much easier and then we can think about an elegant solution to fix it,
instead of hacking it around now.
[~wwei], @[~wilfreds] what do you think? Also please correct me if I missed
something.
> Make sure when there is no allocation in an app, the app state is "Waiting".
> ----------------------------------------------------------------------------
>
> Key: YUNIKORN-402
> URL: https://issues.apache.org/jira/browse/YUNIKORN-402
> Project: Apache YuniKorn
> Issue Type: Sub-task
> Reporter: Kinga Marton
> Assignee: Kinga Marton
> Priority: Major
>
> If there is no allocation for an app, according to
> [http://yunikorn.apache.org/docs/next/design/scheduler_object_states] it's
> status should be waiting instead of running, as mentioned here:
> https://issues.apache.org/jira/browse/YUNIKORN-201?focusedCommentId=17186402&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17186402
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
