[jira] [Commented] (YUNIKORN-3461) Originator can have placeholder flag set

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

    [ https://issues.apache.org/jira/browse/YUNIKORN-3461?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18121413#comment-18121413 ] 

Wilfred Spiegelenburg commented on YUNIKORN-3461:
-------------------------------------------------

Sorry for the late reply: The originator pod is a non placeholder pod by default. The placeholder all get the ownership references set to that specific pod.

If the pod does not have a taskgroup name set it MUST not have the placeholder flag set when we pass it on to the core. That is the first issue. The k8shim should unset the placeholder flag on that task. This does not solve the whole problem.

Second problem is when we mark the pod as the originator: we always hard skip the pod as an originator if it has the placeholder flag set. That could filter out legitimate originator pods that had the placeholder flag "faked" In case someone does set the placeholder label and it has a taskgroup name set then the above change to check both will not catch the problem. The pod will however never be marked as the originator.

We can use the fact that a placeholders MUST only have a single pod as the owner reference. If a pod gets deployed by a controller the controller will set an owner reference. See {{getOwnerReference()}} for what we set.

That would give some logic we can implement: If a pod has the placeholder flag set check the owner references.
 * no owner reference -> unset flag
 * more than one owner reference -> unset flag
 * if the owner reference is not for a {{kind=Pod}} or has {{controller=true}} unset the flag

That would limit the issue as far as we can. The rest we can document.

Hope that clears things up.

> Originator can have placeholder flag set
> ----------------------------------------
>
>                 Key: YUNIKORN-3461
>                 URL: https://issues.apache.org/jira/browse/YUNIKORN-3461
>             Project: Apache YuniKorn
>          Issue Type: Bug
>          Components: shim - kubernetes
>            Reporter: Wilfred Spiegelenburg
>            Assignee: Andy Huang
>            Priority: Major
>
> The {{internal.yunikorn.apache.org/placeholder=true}} can be specified on any pod in the system. If the pod does not have a taskgroup name set it will be rejected by the core when submitted.
> The originator pod, that has the taskgroup definitions, could also have the same label set. This is not correct and will have an undefined behaviour on gang scheduling.
> A pod should not have the placeholder flag if it has the taskgroup definitions included.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

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