Use ContainerPlatform.DEFAULT for selector label generation in KubeMirrorStrategy

[munishchouhan]

This PR will fix #952

[pditommaso]

Would not be better request.platform ?: ContainerPlatform.DEFAULT ?

Also worth adding a test

[munishchouhan]

We have hard-coded null, but a yes condition will be good for future changes.
Sure, I will add the test

[munishchouhan]

done

[pditommaso]

I understand this better now. In reality the request.platform is not related with the platform used to run the container.

The request.platform is used to determine the container platform to be copied, it used null with the meaning all.

The selector determine the platform to be used to run copy job container, however this container is multi-arch i.e. a ARM64 container could copy a AMD64 container.

Ideally this container could run both on AMD64 or ARM64, however I think the deployment env requires at least to be chosen.

@gavinelder what's your take here, considering this job could run on both arch, could the selector be omitted or there's a selected that would allow to run on both ?

[gavinelder]

We have a couple of options here based on the behaviour we want.

Node pools can have more than one label so if you're limited to specifying only one label on the deployment by using the nodeSelector resource we can add an additional label to both the AMD64 / ARM nodes allowing pods to be scheduled on both.

The more flexible but difficult task to implement is to support NodeAffinity rules which work as a node selector but can have more complex logic.

For example.

apiVersion: batch/v1
kind: Job
metadata:
  name: wave-scan
spec:
  template:
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: service  
                operator: In
                values:
                - wave-build-amd
                - wave-build-arm
      containers:
      - name: job-container
        image: my-image
      restartPolicy: Never

Which would change the interface from a string to a list of strings on the config.

[pditommaso]

For consistency I'd do seqera.io/wave-build-any=true instead of seqera.io/wave-supplemental=true

[gavinelder]

@munishchouhan the keys need to be unique as such

'service=wave-build'
'service=wave-build-arm64'

Would not be possible.

seqera.io/wave-build-any=true

It is.

[munishchouhan]

but if we donot set any nodeselector for a pod where we want it to use seqera.io/wave-build-any=true?

[munishchouhan]

another approach #954

[munishchouhan]

please let me know which approach we want to take

1. add node selector in mirror pod
2. keep it null