We have just released Kueue v0.17.0.
For more details, please check the release notes and the documentation.
Highlights:
1. Performance and Scalability: Scheduler throughput is significantly improved for large clusters. Key enhancements include faster TAS snapshotting, alpha support for bulk scheduling with equivalence hashes (
SchedulingEquivalenceHashing), smarter throttling of inadmissible requeues, and reduced memory usage through stripping
managedFields. See
blogpost.
2. Topology-Aware Scheduling: Node hot swap now supports node taints, and TAS snapshotting performance has improved further. The release also introduces multi-layered scheduling for next-generation accelerator hardware, along with
ResourceTransformations support for virtual credit–style resources to enable quota sharing across multiple flavors.
3. MultiKueue: now supports
LeaderWorkerSet and
RayService.. It also introduces alpha support for orchestrated preemption, ensuring that only one worker cluster performs preemption at a time.
4. Expanded AI/ML Integrations: Kueue continues to strengthen its ML ecosystem support with native KubeRay
RayService management, support for Kubeflow Trainer v2.2, and a new
SparkApplication integration available behind a feature gate.
5. KueueViz and Observability: KueueViz adds token-based authentication, resource reporting, and a hierarchical tree view. Observability is also enhanced with new cohort metrics, custom Prometheus labels, and the graduation of
LocalQueue metrics to beta, enabled by default.
6. Scheduling and Fairness: Fair Sharing now favors workloads within nominal quota more predictably. New controls include admission gating with the
kueue.k8s.io/admission-gated-by annotation, and priority boosting via the workload annotation
kueue.x-k8s.io/priority-boost. DRA integration also adds alpha support for extended resources.
7. Core Features Graduated to GA: Several key queueing and cohort-management capabilities are now generally available, including
LendingLimit,
LocalQueueDefaulting,
ObjectRetentionPolicies,
SanitizePodSets, and
HierarchicalCohorts
Thank you to the contributors who reported issues, implemented features, fixed bugs, etc.
Cheers,
Michal