[
https://issues.apache.org/jira/browse/YUNIKORN-742?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17378822#comment-17378822
]
Wilfred Spiegelenburg commented on YUNIKORN-742:
------------------------------------------------
I have read through the KEP and the PRs for the pod deletion cost and there is
no impact on the scheduler. The choice made in the implementation is to define
an annotation on the pod that can be updated.
The annotation is read by the replica set controller. It is processed as part
of the decision which pod to delete. The annotation reading is limited to the
part of the controller pod sorting.
The kubernetes scheduler is not involved at all when it comes to deleting a pod
in the replica set.
This comment in the replica set page does not make sense:
{quote}This works if the application itself controls the down scaling; for
example, the driver pod of a Spark deployment.
{quote}
A spark application is not deployed as a replica set. Especially the driver is
never deployed with multiple replicas and it is just one pod. The executors are
not deployed as a replica set when you run normal spark on kubernetes. The
spark operator is also not using a replica set.
At this point I do not see how to fits in YuniKorn and thus why we should look
at this.
> Add support for pod deletion cost
> ---------------------------------
>
> Key: YUNIKORN-742
> URL: https://issues.apache.org/jira/browse/YUNIKORN-742
> Project: Apache YuniKorn
> Issue Type: Improvement
> Reporter: Holden Karau
> Priority: Minor
>
> Add support for pod deletion costs in the same/similar way as the default
> kubernetes scheduler (FEATURE STATE: Kubernetes v1.21 [alpha] -
> [https://kubernetes.io/docs/concepts/workloads/controllers/replicaset/] )
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
