kubernetes-sigs/kueue v0.14.5

on GitHub

Changes since v0.14.4:

Urgent Upgrade Notes

(No, really, you MUST read this before you upgrade)

• TAS: It supports the Kubeflow TrainJob.

  You should update Kubeflow Trainer to v2.1.0 at least when using Trainer v2. (#7755, @IrvingMg)

Changes by Kind

Bug or Regression

• AdmissionFairSharing: Fix the bug that occasionally a workload may get admitted from a busy LocalQueue,
  bypassing the entry penalties. (#7914, @IrvingMg)

• Fix a bug that an error during workload preemption could leave the scheduler stuck without retrying. (#7818, @olekzabl)

• Fix a bug that the cohort client-go lib is for a Namespaced resource, even though the cohort is a Cluster-scoped resource. (#7802, @tenzen-y)

• Fix integration of manageJobWithoutQueueName and managedJobsNamespaceSelector with JobSet by ensuring that jobSets without a queue are not managed by Kueue if are not selected by the managedJobsNamespaceSelector. (#7762, @MaysaMacedo)

• Fix issue #6711 where an inactive workload could transiently get admitted into a queue. (#7939, @olekzabl)

• Fix the bug that a workload which was deactivated by setting the spec.active=false would not have the
  wl.Status.RequeueState cleared. (#7768, @sohankunkerkar)

• Fix the bug that the kubernetes.io/job-name label was not propagated from the k8s Job to the PodTemplate in
  the Workload object, and later to the pod template in the ProvisioningRequest.

  As a consequence the ClusterAutoscaler could not properly resolve pod affinities referring to that label,
  via podAffinity.requiredDuringSchedulingIgnoredDuringExecution.labelSelector. For example,
  such pod affinities can be used to request ClusterAutoscaler to provision a single node which is large enough
  to accommodate all Pods on a single Node.

  We also introduce the PropagateBatchJobLabelsToWorkload feature gate to disable the new behavior in case of
  complications. (#7613, @yaroslava-serdiuk)

• Fix the race condition which could result that the Kueue scheduler occasionally does not record the reason
  for admission failure of a workload if the workload was modified in the meanwhile by another controller. (#7884, @mbobrovskyi)

• TAS: Fix the requiredDuringSchedulingIgnoredDuringExecution node affinity setting being ignored in topology-aware scheduling. (#7937, @kshalot)