[Feature]: Change nodeAffinity in OLM deployment to not require master or control plane labels on nodes.

[bhpratt]

Suggestion Description

Some installations of managed OpenShift such as Hypershift: https://github.com/openshift/hypershift deploy a hosted control plane that is segmented from the data plane. In these installs, the worker nodes do not have the node-role.kubernetes.io/control-plane or node-role.kubernetes.io/master labels. Because the worker nodes do not have these labels, operators, such as the AMD GPU Operator that include requiredDuringSchedulingIgnoredDuringExecution such as:

spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: node-role.kubernetes.io/control-plane
            operator: Exists
        - matchExpressions:
          - key: node-role.kubernetes.io/master
            operator: Exists

Will not run out of the box. Suggestion is to change the nodeAffinity to preferredDuringSchedulingIgnoredDuringExecution or expose a method for the end user to set their own nodeAffinity to accomodate these types of OpenShift installs.

Operating System

N/A

GPU

N/A

ROCm Component

N/A

[yansun1996]

Hi @bhpratt thanks for the suggestion for deploying AMD GPU operator on a hosted control plane, we will evaluate your PR and get back to you.

[yansun1996]

Hi @bhpratt we have published AMD GPU Operator v1.2.1 into OpenShift OperatorHub as certified operator. The new nodeAffinity you proposed for switching to preferredDuringSchedulingIgnoredDuringExecution has been included in the new v1.2.1 OLM bundle. Please try with the v1.2.1 release. https://catalog.redhat.com/software/containers/amd/gpu-operator-v1/675a43ffba424dccbad17635

Let us know if you have any other question or suggestion.

[bhpratt]

Thanks. I'll try it out!

[bhpratt]

Confirmed in 1.21. Thanks again.

OOC - I didn't see the change in this codebase. I take it there's an internal codebase where the change was made?

      affinity:
        nodeAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
          - preference:
              matchExpressions:
              - key: node-role.kubernetes.io/control-plane
                operator: Exists
            weight: 1
          - preference:
              matchExpressions:
              - key: node-role.kubernetes.io/master
                operator: Exists
            weight: 1

[yansun1996]

Confirmed in 1.21. Thanks again.

OOC - I didn't see the change in this codebase. I take it there's an internal codebase where the change was made?

      affinity:
        nodeAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
          - preference:
              matchExpressions:
              - key: node-role.kubernetes.io/control-plane
                operator: Exists
            weight: 1
          - preference:
              matchExpressions:
              - key: node-role.kubernetes.io/master
                operator: Exists
            weight: 1

Hi @bhpratt

yep we have another workspace to prepare the certification.

As for your PR, it is currently being hold because I need to confirm with our team about how to handle the external PR. Please wait for that.

In addition to this change, is there any other issue that is blocking the deployment on hosted control plane ?