[ANNOUNCE] kueue 0.19.0 is released

[Mykhailo Bobrovskyi]

Hello folks,

We have just released Kueue v0.19.0.

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

Highlights:

1. Topology-Aware Scheduling: TAS continues to scale and mature. Compact assignment encoding now supports mega-scale workloads spanning more than 100k nodes, multi-layer topology support for advanced accelerators graduates to Beta. Workload evaluation performance is significantly improved, particularly for preemption-heavy scenarios; see #13148.
2. MultiKueue: The incremental dispatcher now has configurable step size, giving operators greater control over the number of worker clusters nominated in each dispatching step. New MultiKueue-dedicated metrics provide visibility into MultiKueue dispatch and admission; see #11132.

3. WAS & TAS  Integration: Workload-Aware Scheduling can now use the Kubernetes Scheduler Library to perform TAS node-domain fit checks. This new alpha integration brings Kueue’s workload admission decisions closer to Kubernetes scheduling semantics; see #8871.
4. Dynamic Resource Allocation: The extended-resource and partitionable-device integrations graduate to Beta. Kueue also adds consumable-capacity for devices supporting multiple allocations and multi-counter tracking for partitionable devices.
5. Elastic Jobs: Elastic workloads receive broad reliability fixes. MultiKueue integration is improved by propagating elastic RayCluster scaling changes from the management cluster to the admitting worker cluster.

6. ConcurrentAdmission: Concurrent admission now supports preemption control through preemption gates, ensuring that only one admission Variant can issue preemptions at a time. Variants are also kept synchronized when the ResourceFlavors of their ClusterQueue change.
7. Observability: New diagnostic reasons and messages for the QuotaReserved=False condition make it easier to understand why a Workload remains unadmitted; this is an alpha feature. Additionally, CustomMetricLabels enables selected metrics to be extended with labels sourced from Workload objects. KueueViz receives further security, reliability, performance, and usability improvements.
8. AI scan-based hardening: The project addressed more than half of the bugs and vulnerabilities identified by the AI-assisted scan tracked in #12573. The hardening effort continues, with the goal of resolving the remaining findings.
9. Smarter defaults: WaitForPodsReady is now enabled by default, with a 30-minute timeout. Kueue’s out-of-the-box Kubernetes client limits increase to 300 QPS. Default controller concurrency increases to 10 for Workloads and 5 each for LocalQueues and ClusterQueues.

Thank you to the contributors who reported issues, implemented features, fixed bugs, etc.

Cheers,

Mykhailo on behalf of the Kueue Release Team

To unsubscribe from this group and stop receiving emails from it, send an email to wg-batch+u...@kubernetes.io.