Service selector conflicts with LabelTransformer in ROCm TensorFlow kustomize manifests

[coderabbitai]

Problem Description

The Service manifest in jupyter/rocm/tensorflow/ubi9-python-3.12/kustomize/base/service.yaml has hardcoded app: notebook labels in both metadata.labels and spec.selector, while the kustomization layer adds an additional app: jupyter-rocm-tensorflow-ubi9-python-3-12 label to both objects and their selectors (includeSelectors: true).

After kustomize render, the Service selector will demand both labels (app: notebook AND app: jupyter-rocm-tensorflow-ubi9-python-3-12), yet the pod template in the StatefulSet only receives the generated label (it has no app: notebook). This results in an empty Endpoints list and 503 errors when trying to access the notebook service.

Expected Behavior

The Service should be able to route traffic to the notebook pods successfully.

Proposed Solution

Update the service's metadata.labels and spec.selector to rely on kustomize to inject the canonical app label:

metadata:
  name: notebook
  # Let kustomize inject the canonical app label;
  # keep this block minimal to avoid selector divergence.
  labels: {}
spec:
  type: ClusterIP
  ports:
    - port: 8888
      protocol: TCP
      targetPort: notebook-port
  # Rely on kustomize to fill in the selector.
  selector: {}

References

• PR: RHOAIENG-27434: Create ROCm Tensorflow Python 3.12 Image #1259
• Comment: RHOAIENG-27434: Create ROCm Tensorflow Python 3.12 Image #1259 (comment)
• Related issues: Service/Pod label mismatch in ROCm PyTorch kustomize manifests causes connectivity issues #1251 (same problem in PyTorch), StatefulSet selector and template labels configuration issue in jupyter/rocm/tensorflow #1264 (StatefulSet issue in same component)
• Reported by: @jiridanek

[coderabbitai]

Closing as duplicate of #1236 which now covers both StatefulSet and Service selector issues across all Jupyter images systematically. The consolidated issue provides a more comprehensive solution approach.