[
https://issues.apache.org/jira/browse/YUNIKORN-971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17479124#comment-17479124
]
Weiwei Yang commented on YUNIKORN-971:
--------------------------------------
hi [~ccondit], thank you for getting this done and thank you [~pbacsko] for the
review.
Sorry I just got some time looking at the details, my major concern is: we are
tracking all nodes' info in yunikorn, and that runs in parallel with the
default scheduler. Potentially it may happen that YK thinks the pod should be
allocated on node A, but when we pass that to the default scheduler, it fails.
I am worried about the following issue:
if a pod is unable to be allocated on A, in today's implementation the pod will
be left unschedulable. this is not good, it should be retried on a different
node. YK or the default scheduler doesn't have this issue because as long as
they decide to allocate the pod onto a node, it will start to do the binding
steps. but here, even YK-plugin decides to allocate on A, the default scheduler
will still need to run its own validations before binding the pod, so the above
issue could happen.
I want to understand more about the reason for implementing the Filter plugin.
I understand YK can select its own priority list of nodes, but how that is
different than the default scheduler? Ideally, we can add some customized Score
plugin
([https://kubernetes.io/docs/concepts/scheduling-eviction/scheduling-framework/#scoring)]
to control the behavior, why we need to do this by implementing the Filter
plugin?
Another thing, [~ccondit], can you put your design doc in a google doc, I have
some comments to add. And once the doc is settled, we need to publish it to the
website.
> Implement YuniKorn as a Kubernetes Scheduler Plugin
> ---------------------------------------------------
>
> Key: YUNIKORN-971
> URL: https://issues.apache.org/jira/browse/YUNIKORN-971
> Project: Apache YuniKorn
> Issue Type: New Feature
> Components: shim - kubernetes
> Reporter: Craig Condit
> Assignee: Craig Condit
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.0.0
>
> Attachments: YuniKorn as K8S Scheduler Plugin.pdf
>
>
> As of Kubernetes 1.19, there is a new [scheduling
> framework|https://kubernetes.io/docs/concepts/scheduling-eviction/scheduling-framework/]
> available to scheduler implementations. This exposes all the functionality
> of the default Kubernetes scheduler, but allows plugins to be added which
> augment and extend the default scheduler functionality.
> We should explore implementing YuniKorn using this new framework. See
> attached design document for rationale and approach.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
