[
https://issues.apache.org/jira/browse/YUNIKORN-2421?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17818149#comment-17818149
]
Niels Geuze commented on YUNIKORN-2421:
---------------------------------------
I tried to reproduce this but I seem to be unable to, so far I've only seen it
in our production cluster and as a workaround I had disabled reservations. I've
reenabled them now if it occurs again I'll update. The strange part is that
during the debugging I found that reservations should be skipped since we are
not using Gang Scheduling anymore (and as far as I could tell it is skipped
when there are no task groups defined)
> Do not make reservations for pods that have Pending Volumes
> -----------------------------------------------------------
>
> Key: YUNIKORN-2421
> URL: https://issues.apache.org/jira/browse/YUNIKORN-2421
> Project: Apache YuniKorn
> Issue Type: Bug
> Reporter: Niels Geuze
> Priority: Major
>
> I am using Persistent Volume Claims which use a storageclass that has a
> topology constraint (allowedTopologies) to it (I do not know whether this
> constraint is relevant). Today I had several times that for some reason the
> Persistent Volume Claim remained in the Pending state but Yunikorn still
> decided to make a reservation for the pod on a node. This prevented any other
> pods from being scheduled on said node util the PVC was Bound and the pod
> could start or until the pod was deleted.
> I think Yunikorn should check whether all mounted PVCs are Bound and only
> then try to make a reservation to prevent locking nodes that could do other
> tasks, or alternatively release the reservation when the pod cannot be
> allocated.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
