K8s ConfigMap missing in k8shim repository

[Ma...@swisscom.com]

Dear Yunikorn developers

I have tried to deploy Yunikorn in Scheduler plugin mode. Unfortunately, we cannot use helm in our environment, so I fell back to the provided manual steps in the documentation<https://yunikorn.apache.org/docs/developer_guide/deployment>.
I have some questions / issues that prevent me from using your great project:


  1.  In the plugin.yaml<https://github.com/apache/yunikorn-k8shim/blob/master/deployments/scheduler/plugin.yaml#L127> file there is a reference to a ConfigMap called “yunikorn-scheduler-plugin-configs”. Unfortunately, this ConfigMap is completely missing from both the documentation and the repository. Can you somehow provide this ConfigMap, or which settings need to be set in this to get it to work?
  2.  In the documentation it is also mentioned that “Plugin mode is currently very new and has therefore not yet reached the maturity level of standard mode”. Since our cluster has mixed workloads, we cannot run Yunikorn in standard mode. Could you give us some testimony / information in terms of estimated stability of the plugin mode vs the standard mode?

Thanks for your answers. Please let me know if you have any further questions for me to clarify.

Marc Singer

Re: K8s ConfigMap missing in k8shim repository

[Craig Condit <cc...@apache.org>]

Hi Marc,

You can use helm locally to generate the templates for YuniKorn and save them to local files which can be applied. Something like this should work:

    helm repo add yunikorn https://apache.github.io/yunikorn-release
    helm template --namespace yunikorn yunikorn/yunikorn --version 1.2.0 --set enableSchedulerPlugin=true > out.yaml

Add any helm variables as necessary for your environment and deploy the resulting resources. The plugin.yaml file you referenced is not accurate, and won’t work.

YuniKorn can be used with mixed workloads even in standard mode, but there’s slightly better compatibility in plugin mode, as the same operators and plugins the default scheduler uses for making decisions are used. The main limitation currently of the plugin mode is that it doesn’t play nicely with autoscalers — there are situations where unwarranted scale-up can occur as pods temporarily go into Unschedulable state until they are evaluated by YuniKorn. This will be addressed in an upcoming release when we are able to make use of some new plugin functionality in later Kubernetes releases.

Craig



> On Feb 8, 2023, at 9:05 AM, Marc.Singer@swisscom.com wrote:
> 
> Dear Yunikorn developers
> 
> I have tried to deploy Yunikorn in Scheduler plugin mode. Unfortunately, we cannot use helm in our environment, so I fell back to the provided manual steps in the documentation<https://yunikorn.apache.org/docs/developer_guide/deployment>.
> I have some questions / issues that prevent me from using your great project:
> 
> 
>  1.  In the plugin.yaml<https://github.com/apache/yunikorn-k8shim/blob/master/deployments/scheduler/plugin.yaml#L127> file there is a reference to a ConfigMap called “yunikorn-scheduler-plugin-configs”. Unfortunately, this ConfigMap is completely missing from both the documentation and the repository. Can you somehow provide this ConfigMap, or which settings need to be set in this to get it to work?
>  2.  In the documentation it is also mentioned that “Plugin mode is currently very new and has therefore not yet reached the maturity level of standard mode”. Since our cluster has mixed workloads, we cannot run Yunikorn in standard mode. Could you give us some testimony / information in terms of estimated stability of the plugin mode vs the standard mode?
> 
> Thanks for your answers. Please let me know if you have any further questions for me to clarify.
> 
> Marc Singer


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