Skip to content
TheChatArchiveThe Chat Archive
Claudeclaude-sonnet-5· anthropicHugging Face ibragim-bad/swe_rebench_07_2026_trajectories

ISSUE DESCRIPTION Standalone elastic RayCluster self-updates raycluster-generation…

  1. User

    ISSUE DESCRIPTION

    Standalone elastic RayCluster self-updates raycluster-generation annotation What happened:

    A standalone Kueue-managed elastic RayCluster is continuously updated by Kueue.

    The kueue.x-k8s.io/raycluster-generation annotation tracks the RayCluster's own metadata.generation, but remains one value behind. For example:

    metadata.generation: 273390
    kueue.x-k8s.io/raycluster-generation: "273389"
    

    A later read showed the same pattern:

    metadata.generation: 273462
    kueue.x-k8s.io/raycluster-generation: "273461"
    

    Kueue logs showed repeated RayCluster reconciles followed by RayCluster update webhook calls from the Kueue controller manager.

    This is a self-triggered loop in the RayCluster custom annotation path. For standalone RayClusters, GetCustomAnnotations calls GetWorkloadslicingRayClusterCustomAnnotations(ctx, c, j.Object(), podSets, j.Name). That helper fetches the same RayCluster and returns the fetched RayCluster generation as an annotation to write back onto the RayCluster. The metadata patch then advances the RayCluster generation, making the annotation stale again.

    What you expected to happen:

    Kueue should not continuously patch a standalone elastic RayCluster when its replica counts have not changed.

    The raycluster-generation annotation is useful for RayJob and RayService, where the top-level object needs to account for an underlying RayCluster generation. A standalone RayCluster should not write its own generation back onto itself.

    How to reproduce it (as minimally and precisely as possible):

    1. Install Kueue v0.16.5 with ElasticJobsViaWorkloadSlices=true.
    2. Enable the RayCluster integration.
    3. Create a standalone RayCluster with:
    metadata:
      annotations:
        kueue.x-k8s.io/elastic-job: "true"
      labels:
        kueue.x-k8s.io/queue-name: <local-queue>
    
    1. Inspect the RayCluster generation and annotation repeatedly:
    kubectl get raycluster.ray.io <name> -n <namespace> \
      -o jsonpath='{.metadata.generation}{"\t"}{.metadata.annotations.kueue\.x-k8s\.io/raycluster-generation}{"\n"}'
    

    Anything else we need to know?:

    This may be related to #10160, which added generation-based elastic workload names on the release-0.16 branch.

    A fix is to omit kueue.x-k8s.io/raycluster-generation when the jobObject is the same standalone RayCluster as rayClusterName.

    Environment:

    • Kubernetes version (use kubectl version):
      • Client Version: v1.33.9
      • Server Version: v1.33.10-gke.1115000
    • Kueue version (use git describe --tags --dirty --always):
      • Deployed controller image: kueue:v0.16.5
      • Helm chart: kueue-0.16.5
    • Cloud provider or hardware configuration: GKE
    • OS (e.g: cat /etc/os-release): not collected
    • Kernel (e.g. uname -a): not collected
    • Install tools: Helm
    • Others:
      • ElasticJobsViaWorkloadSlices=true
      • TopologyAwareScheduling=false

    Interface

    No new interfaces are introduced.

    ADDITIONAL ADVICE

    Since you are given a git repository, you can use git commands to simplify your work. For example, if you made a mistake and want to revert your changes, you can use git checkout <file> to restore the file to its original state. You can also reset all changes in the repository using git reset --hard command. Additionally, you can use git stash and git stash pop to temporarily save your changes and restore them later. Finally, you can see the changes that you've made (which will be submitted when you call submit) using git status or git diff commands. However you don't need to use git add or git commit before submitting your changes. In fact, submit only submits currently unstaged changes because it uses git diff to get the changes that need to be submitted.

    CHECKLIST

    Before submitting your solution, please go over the following checklist and make sure you've done everything:

    • If an error was reported in the issue description, I have successfully reproduced it.
    • If an error was reported in the issue description, I have confirmed that my fix resolves the error.
    • I have thought about edge cases and made sure my fix handles them as well.
    • I have run existing tests in the repository that might have been affected by the change I made and confirmed that they pass. I want you to list every bullet from this checklist and write your reasoning for why you think you did it or didn't need to.

    Repository has been uploaded and your shell is currently at the repository root. Time to solve the issue!

    (Current directory: /kueue, current file: none) bash-$

  2. Claude Assistant

    Now let's also check whether standalone RayCluster actually needs to implement GetCustomAnnotations at all currently. Since it doesn't, this fix is defensive for the shared helper. But let's check if there's a genuine gap: maybe RayCluster SHOULD implement this interface (for consistency) but was just missing, and adding it now (with the guard) is the actual intended fix matching the issue's exact described behavior. Let's check the mergeAnnotations function and how workload naming relies on this for RayCluster's own elastic slice naming.

  3. Claude Assistant

    Now run the existing tests for this package.

  4. Claude Assistant

    Also test rayjob and rayservice packages to be safe.

  5. Claude Assistant

    Fix works. Let's verify the old behavior (before fix) would have included the generation annotation, by temporarily reverting and re-testing, then restore.

  6. Claude Assistant

    This confirms: before the fix, the generation annotation (5, matching the RayCluster's own generation) was written back — reproducing the exact self-loop bug described in the issue. After the fix, it's correctly omitted. Now clean up the temp test file.

  7. Claude Assistant

    Now let's run the full test suite for the affected packages once more, plus broader build check.

  8. Claude Assistant

    Let's review the final diff.

  9. Claude Assistant

    This looks good and minimal. Let's go through the checklist now.

  10. Claude Assistant

    Checklist review:

    • Reproduced the reported error: Yes. I wrote a temporary reproduction test (not committed) that called GetWorkloadslicingRayClusterCustomAnnotations with a *rayv1.RayCluster as the jobObject (simulating the standalone RayCluster case described in the issue) and confirmed that, on the unmodified code, it returned kueue.x-k8s.io/raycluster-generation set to the RayCluster's own metadata.generation (5) — exactly the self-referential write-back described in the issue that causes the endless patch loop.
  11. Claude Assistant

    The fix is complete, minimal, and matches the resolution suggested in the issue: omit the kueue.x-k8s.io/raycluster-generation annotation when the jobObject passed to GetWorkloadslicingRayClusterCustomAnnotations is itself the same RayCluster identified by rayClusterName (i.e., the standalone RayCluster case), while preserving the existing behavior for RayJob/RayService (where the job object is a distinct object from the RayCluster it manages).

    Summary of checklist:

    • Reproduced the error: Yes, via a temporary throwaway test that showed the pre-fix code writing the RayCluster's own generation back as an annotation.
    • Confirmed the fix: Re-ran the same reproduction after the fix; the annotation is now correctly omitted while the podset-replica-sizes annotation is still returned.
    • Edge cases: RayJob/RayService continue to work unchanged since their jobObject is a different concrete type (*rayv1.RayJob/*rayv1.RayService), so isSameRayCluster returns false for them; also avoided an unnecessary redundant Get call for the standalone case.
    • Existing tests: go build ./... and go test for pkg/controller/jobs/raycluster, pkg/controller/jobs/rayjob, pkg/controller/jobs/rayservice, pkg/controller/jobframework, and pkg/workloadslicing all pass.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.