support use model-initializer with "no-engine" when modelUri set

[cnfatal]

For some historical reason, I need serving some custom protocol models(image). I'm not able to Building a new inference server for these models, so I have to serve them directly with the annotation seldon.io/no-engine: "true".

Then I cannot use storageInitializer even though the modelUri set, the init container not injected.

As an alt, I set the init container in componentSpecs by hand, but it's not what I want.

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: sklearn-iris
spec:
  predictors:
    - name: default
      replicas: 1
      annotations:
        seldon.io/no-engine: "true"
      graph:
        name: classifier
        type: MODEL
        # modelUri: s3://sklearn/iris
      componentSpecs:
        - spec:
            containers:
              - name: classifier
                image: some-model-image:1.13
                securityContext:
                  privileged: true
                ports:
                  - containerPort: 8080
                    protocol: TCP
                    name: http
                volumeMounts:
                  - mountPath: /mnt/models
                    name: classifier-provision-location
                    readOnly: true
            volumes:
              - name: classifier-provision-location
                emptyDir: {}
            initContainers:
              - name: classifier-model-initializer
                image: seldonio/rclone-storage-initializer:1.15.0-dev
                imagePullPolicy: IfNotPresent
                args:
                  - "s3://sklearn/iris"
                  - "/mnt/models"
                volumeMounts:
                  - mountPath: /mnt/models
                    name: classifier-provision-location
                envFrom:
                  - secretRef:
                      name: seldon-init-container-secret

I'm glad to implement this feature if needed

[cnfatal]

There are some cases：

Case 0:

I have several images of embedded models that implement sekldon's Custom inference servers, But there are too many images to write each of them into the seldon-config configmap.

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: sklearn-iris
spec:
  predictors:
    - name: default
      replicas: 1
      graph:
        name: classifier
        # implementation: null
        modelUri: gs://seldon-models/v1.15.0-dev/sklearn/iris
      componentSpecs:
        - spec:
            containers:
              - name: classifier
                image: model-server-allinone:latest

Case 1(my situation):

In Case 0, some images not implement Custom inference servers,so I must serve them directly with seldon.io/no-engine: "true". It's simple like a web app.

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: sklearn-iris
spec:
  predictors:
    - name: default
      replicas: 1
      annotations:
         seldon.io/no-engine: "true"
      graph:
        name: classifier
        modelUri: gs://seldon-models/v1.15.0-dev/sklearn/iris
        endpoint:
          type: REST
          httpPort: 80
      componentSpecs:
        - spec:
            containers:
              - name: classifier
                image: nginx:1.13
                securityContext:
                  privileged: true
                ports:
                  - containerPort: 80
                    protocol: TCP
                    name: http

In both cases I need to use rclone to download model files from some storage server. But in these cases the initialize container is not injection when modelUri is set.

[ukclivecox]

Closing also please try this in v2