INFRA Signal 147
Kubernetes v1.37: Advancing Workload-Aware Scheduling
↗
Illustration only Photo by Teng Yuhong on Unsplash
Kubernetes v1.37 graduates Workload, PodGroup, gang scheduling, workload-aware preemption, and shared DRA ResourceClaims to Beta, introduces the CompositePodGroup API for hierarchical scheduling, and makes PodGroup a first-class queueing citizen with mutable minCount.
For engineers running AI/ML or batch workloads, gang scheduling and workload-aware preemption are now one step from General Availability, making them viable for production use. The new CompositePodGroup API and controller integration tooling reduce the need for custom scheduling extensions, while mutable minCount enables elastic workloads to resize without disrupting running Pods.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Workload, PodGroup, gang scheduling, and workload-aware preemption APIs graduate to Beta at v1beta1, replacing v1alpha2 with breaking changes to the disruptionMode structure.
The new CompositePodGroup API enables multi-level topology constraints and gang scheduling for heterogeneous Pod groups managed by JobSet or LeaderWorkerSet.
PodGroup becomes a first-class queueing object, minCount is now mutable, and default preemption now respects PodGroup disruptionMode fields.
THE CLUSTER