Some V2 pre-packaged servers do not allow env var override

[adriangonz]

Describe the bug

The V2 pre-packaged servers, at least the ones backed by MLServer, do not allow overriding certain environment variables (e.g. MLSERVER_MODEL_URI). Instead, they will always force the default values.

This can become a blocker for certain scenarios (e.g. multi-model serving), where we want to remove the default value to let each model decide on their own.

To reproduce

1. Deploy a SeldonDeployment manifest like the one below:

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: my-model
spec:
  protocol: v2
  predictors:
  - name: default
    graph:
      name: model
      implementation: SKLEARN_SERVER
      modelUri: gs://seldon-models/mlserver/mms3
    componentSpecs:
      - spec:
          containers:
          - name: model
            env:
            - name: MLSERVER_MODEL_URI
              value: ""

2. Inspect the pod.
3. You will notice that the MLSERVER_MODEL_URI env var is still set to /mnt/models.

Expected behaviour

The operator should merge the env vars provided through the SeldonDeployment manifest with the default values, to allow the user to override their value.

Environment

Seldon Core 1.15.0-dev

Model Details

• Images of your model: [Output of: kubectl get seldondeployment -n <yourmodelnamespace> <seldondepname> -o yaml | grep image: where <yourmodelnamespace>]
• Logs of your model: [You can get the logs of your model by running kubectl logs -n <yourmodelnamespace> <seldonpodname> <container>]