github kubernetes-sigs/kueue v0.9.4

22 hours ago

Changes since v0.9.3:

Changes by Kind

Bug or Regression

  • Add missing external types to apply configurations (#4202, @astefanutti)

  • Disable the StatefulSet webhook in the kube-system and kueue-system namespaces by default.
    This aligns the default StatefulSet webhook configuration with the Pod and Deployment configurations. (#4161, @@dgrove-oss)

  • Fix a bug is incorrect field path in inadmissible reasons and messages when Pod resources requests do not satisfy LimitRange constraints. (#4290, @tenzen-y)

  • Fix a bug is incorrect field path in inadmissible reasons and messages when container requests exceed limits (#4246, @tenzen-y)

  • Fix a bug that allowed unsupported changes to some PodSpec fields which were resulting in the StatefulSet getting stuck on Pods with schedulingGates.

    The validation blocks mutating the following Pod spec fields: nodeSelector, affinity, tolerations, runtimeClassName, priority, topologySpreadConstraints, overhead, resourceClaims, plus container (and init container) fields: ports and resources.requests.

    Mutating other fields, such as container image, command or args, remains allowed and supported. (#4154, @mbobrovskyi)

  • Fix a bug that doesn't allow Kueue to delete Pods after a StatefulSet is deleted. (#4206, @mbobrovskyi)

  • Fix a bug that prevented tracking some of the controller-runtime metrics in Prometheus. (#4227, @tenzen-y)

  • Fix a bug truncating AdmissionCheck condition message at 1024 characters when creation of the associated ProvisioningRequest or PodTemplate fails.
    Instead, use the 32*1024 characters limit as for condition messages. (#4195, @mbobrovskyi)

  • Fix the bug that prevented Kueue from updating the AdmissionCheck state in the Workload status on a ProvisioningRequest creation error. (#4118, @mbobrovskyi)

  • Helm: Fix the unspecified LeaderElection Role and Rolebinding namespaces (#4386, @eric-higgins-ai)

  • MultiKueue: Do not update the status of the Job on the management cluster while the Job is suspended. This is updated for jobs represented by JobSet, Kubeflow Jobs and MPIJob. (#4085, @IrvingMg)

  • Propagate the top-level setting of the kueue.x-k8s.io/priority-class label to the PodTemplate for
    Deployments and StatefulSets. This way the Workload Priority class is no longer ignored by the workloads. (#4036, @Abirdcfly)

  • TAS: Fix a bug that unschedulable nodes (".spec.unschedulable=true") are counted as allocatable capacities (#4209, @tenzen-y)

  • TAS: Fixed a bug that allows to create a JobSet with both kueue.x-k8s.io/podset-required-topology and kueue.x-k8s.io/podset-preferred-topology annotations set on the PodTemplate. (#4156, @mbobrovskyi)

Other (Cleanup or Flake)

  • Renamed Log key from "attemptCount" to "schedulingCycleCount". This key tracks how many scheduling cycles we have done since starting Kueue. (#4241, @tenzen-y)

Don't miss a new kueue release

NewReleases is sending notifications on new releases.