[ANNOUNCE] Kueue v0.15.0 is released

[Michał Woźniak]

Hello kubefolks, we have released Kueue v0.15.0.

• API Version Graduation  The v1beta2 API is now exposed for CRD serving and config APIs. While v1beta1 remains supported for storage in this release, migration is highly recommended. Storage will move to v1beta2 in v0.16, and v1beta1 support will be discontinued in v0.17. This is a major milestone towards V1!
• Topology-Aware Scheduling  Topology-Aware Scheduling (TAS) is now supported for the Kubeflow Trainer v2. We also introduced Balanced Placement (alpha) in TAS to optimize network performance for multi-layer accelerators. Finally, v1beta2 will allow TAS to support workloads up to 60k in 0.16 - stay tuned!
• MultiKueue Capabilities Expanded 🌐 MultiKueue now supports Topology Aware Scheduling (TAS) and ProvisioningRequest integration for complex multi-cluster scenarios. Additionally, ClusterProfile integration offers a standardized, vendor-neutral interface for cluster information using cloud-provider dedicated plugins. We have hardened security with tighter kubeconfig validation and improved support for role-sharing as a management and worker cluster.
• Simplified Pod Integration  Pod integration is now automatically enabled when using LeaderWorkerSet, StatefulSet, or Deployment frameworks.
• DRA Integration Update  The Dynamic Resource Allocation (DRA) integration has been updated to use DRA API v1 (migrating from v1beta2). Please note that enabling this feature now requires Kubernetes 1.34+.
• Feature Graduations to Beta  Several key features have graduated to Beta and are enabled by default:
  • AdmissionFairSharing: Fair sharing for ClusterQueues among LocalQueues based on historical usage.
  • MultiKueueBatchJobWithManagedBy: Automatic status synchronization for batch/Jobs.
  • MultiKueueAdaptersForCustomJobs: MultiKueue support for external Jobs.
  • ManagedJobsNamespaceSelectorAlwaysRespected: Generalization of the selector in the config API.
• PodFailureRecovery (Alpha)  Introducing a new Pod failure controller to gracefully handle Pods stuck in a terminating state due to Node failure. This feature activates only for Pods annotated with kueue.x-k8s.io/safe-to-forcefully-terminate.

For more details, please check the release notes and the documentation.

As always, feel free to share your feedback or join the discussions, either via Kueue issues or the wg-batch.

Cheers,
Michał