vllm-project / vllm-project/production-stack
bug: The VLLMRuntime controller fails on every reconciliation when running on a cluster that does not have KEDA installed
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 2.6k
- Forks
- 503
- Avg merge
- 4d 17h
- Merged PRs (30d)
- 8
Description
Describe the bug
The VLLMRuntime controller fails on every reconciliation when running on a cluster that does not have KEDA installed, even if no VLLMRuntime resource enables autoscaling. This makes the operator unusable on clusters without KEDA.
To Reproduce
Environment
- vLLM Production Stack Operator
- Cluster without KEDA (no
keda.shAPI group registered)
Steps to Reproduce
- Install the operator on a cluster without KEDA
- Create a VLLMRuntime resource without
autoscalingConfig(or withautoscalingConfig.enabled: false) - Observe the controller logs
Expected behavior
The VLLMRuntime reconciles successfully. Since autoscaling is not requested, the operator should not require KEDA to be present.
Additional context
Root cause:
In operator/internal/controller/vllmruntime_controller.go:360-370, when autoscalingConfig is nil or disabled, the else branch attempts to delete a ScaledObject as a cleanup measure:
The guard !errors.IsNotFound(err) only catches the case where the ScaledObject resource does not exist (i.e., the keda.sh API group is registered, but no ScaledObject with that name is found). It does not catch the case where the keda.sh API group itself is not registered on the cluster. In that scenario, the API server returns a different error (e.g., "the server could not find the requested resource") that does not satisfy errors.IsNotFound(), causing the reconciliation to fail.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start in operator/internal/controller/vllmruntime_controller.go around lines 360-370 and reproduce the reconciliation on a cluster without the keda.sh API group, using a VLLMRuntime with autoscaling omitted or disabled. Confirm the controller behavior for that scenario and verify that reconciliation succeeds without KEDA while autoscaling remains unrequested.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- go, kubernetes
- Domain
- infrastructure
- Issue type
- Bug
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Activity status
- Quiet
- Clarity
- Clearly specified
- Newbie friendliness
- 72/100