[jira] [Commented] (YUNIKORN-586) Enhance placeholder cleanup on timeout

["Wilfred Spiegelenburg (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/YUNIKORN-586?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17304625#comment-17304625 ] 

Wilfred Spiegelenburg commented on YUNIKORN-586:
------------------------------------------------

Cleanup can happen at two points in time:
 # some placeholders have been allocated and some have not
 # all placeholders are allocated and real allocations are scheduled

The first case is not an issue. All allocated placeholders are released, the left over asks are released. There can be no replacements in flight yet.

The second case cleanup is going to be be a problem. When we schedule real allocations we have mapped that real allocation to a placeholder. The swap cannot happen until the placeholder is released. We may not look at that placeholder again for other real allocations and have marked a placeholder as released. The real allocation is linked to this placeholder. When the shim confirms the release the swap happens. Releasing the placeholder can take some time. We depend on the response from the shim and cluster underneath for that to happen.

If the timeout for the placeholder is triggered while a replacement is ongoing we should let that replacement finish. We already have requested a release and should not release that placeholder again.

> Enhance placeholder cleanup on timeout
> --------------------------------------
>
>                 Key: YUNIKORN-586
>                 URL: https://issues.apache.org/jira/browse/YUNIKORN-586
>             Project: Apache YuniKorn
>          Issue Type: Bug
>          Components: core - scheduler
>            Reporter: Wilfred Spiegelenburg
>            Assignee: Wilfred Spiegelenburg
>            Priority: Major
>
> When we cleanup placeholders on timeout we currently release all placeholders that have been allocated without filtering.
> This does not seem correct and does not account for the case that there could be replacements that are in flight. We must filter out the replacements that are in flight when we clean up.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@yunikorn.apache.org
For additional commands, e-mail: issues-help@yunikorn.apache.org