ability to specify init container on per deployment basis

[RafalSkolasinski]

Currently init containers can only be specified globally. It would be good if one can specify init container on per deployment basis.

[ukclivecox]

You can add your own initContainer in the connected PodSpec. But if you are using a prepackaged server you will still get its initContainer added by default which may conflict.

[RafalSkolasinski]

Actually, tested it just now and I only got one init container with prepackaged server. This is most probably because name matched the pattern. The example I came up with is

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: example-test
spec:
  name: iris
  predictors:
  - componentSpecs:
    - spec:
        containers:
        - name: classifier
          volumeMounts:
          - mountPath: /mnt/models
            name: classifier-provision-location
            readOnly: true
    
        initContainers:
        - name: classifier-model-initializer
          image: rafalskolasinski/storage-initializer:v0.4.0
          imagePullPolicy: IfNotPresent
          args:
          - s3://sklearn/iris
          - /mnt/models
    
          envFrom:
          - secretRef:
              name: seldon-init-container-secret
    
          terminationMessagePath: /dev/termination-log
          terminationMessagePolicy: File
    
          volumeMounts:
          - mountPath: /mnt/models
            name: classifier-provision-location
    
        volumes:
        - emptyDir: {}
          name: classifier-provision-location
    
    graph:
      children: []
      implementation: SKLEARN_SERVER
      modelUri: s3://sklearn/iris
      name: classifier
    name: default
    replicas: 1

(which only customises the secret and image)

[RafalSkolasinski]

Adding above example as part of #2055 should solve this issue.

[RafalSkolasinski]

Thinking about it more, we should consider this issue close only when it will be possible to specify image used for initContainer per deployment in a more elegant fashion than above.

[RafalSkolasinski]

To be precise, I would like to allow users to be able to specify a following

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: custom-init-image
spec:
  name: iris
  predictors:
  - componentSpecs:
    - spec:
        initContainers:
        - name: classifier-model-initializer
          image: seldonio/storage-initializer:v1.6.0
    graph:
      implementation: SKLEARN_SERVER
      modelUri: gs://seldon-models/sklearn/iris
      name: classifier
    name: default
    replicas: 1

With the assumption that all valid init container images provided these way would take two arguments:

• source
• destination

[RafalSkolasinski]

This should in principle work like this but currently operator fails to create a proper deployments: #2821

[RafalSkolasinski]

closed by #2937