[
https://issues.apache.org/jira/browse/YUNIKORN-2664?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17852380#comment-17852380
]
Craig Condit commented on YUNIKORN-2664:
----------------------------------------
This is a hard Won’t Do. Maintaining consistent internal state between multiple
schedulers of the complexity of YuniKorn to support hot standby is nearly
impossible, especially when the Kubernetes API server is asynchronous in
nature. Instead, we have opted to make YuniKorn completely stateless externally
so that restarts are fast and consistent. Even in very large clusters,
restarting a new instance via a Deployment typically takes only seconds.
> Running YuniKorn as leader-elected controller with multiple replicas
> --------------------------------------------------------------------
>
> Key: YUNIKORN-2664
> URL: https://issues.apache.org/jira/browse/YUNIKORN-2664
> Project: Apache YuniKorn
> Issue Type: Wish
> Components: shim - kubernetes
> Reporter: Volodymyr Kot
> Priority: Major
>
> Hey, I noticed that by default YuniKorn is run as a Deployment with a single
> replica:
> [https://github.com/apache/yunikorn-release/blob/aa9a2939eed81fc74fbbf7afbc0fe60c5aa0acd0/helm-charts/yunikorn/templates/deployment.yaml#L31]
> and leader election is disabled in scheduler configuration:
> [https://github.com/apache/yunikorn-k8shim/blob/36111c41d97658e168e640c284fe8d71921883b4/conf/scheduler-config.yaml#L20]
>
> Is there anything about the architecture of YuniKorn that makes this
> hard/impossible to do? Or would you be open to a PR that adds ability to run
> with multiple replicas and leader election?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
