Kubernetes 1.36 splits Workload and PodGroup APIs for batch and AI scheduling

Kubernetes v1.36 ships the second iteration of workload-aware scheduling, announced on the project blog May 13, 2026. The shape of the API has changed since v1.35.

A clean split

v1.35 introduced the Workload API and basic gang scheduling on a Pod-based framework. v1.36 moves both APIs to scheduling.k8s.io/v1alpha2 and splits concerns: the Workload acts as a static template containing podGroupTemplates, while the new PodGroup object carries the runtime schedulingPolicy and status.conditions. The Pod field workloadRef is replaced with schedulingGroup, which references a PodGroup directly so the scheduler can read scheduling state without parsing the Workload.

What landed under the new shape

• A new PodGroup scheduling cycle in kube-scheduler for atomic workload processing
• First iterations of topology-aware scheduling and workload-aware preemption
• ResourceClaim support for workloads, extending Dynamic Resource Allocation to entire PodGroups
• The first phase of integration between the Job controller and the new API

Still alpha

Everything in v1alpha2 is alpha. Don’t migrate production gang-scheduling off Kueue, Volcano, or YuniKorn yet. But the upstream foundations for the same patterns now exist, and the API shape will keep moving until beta.

Source: Kubernetes v1.36: Advancing Workload-Aware Scheduling — May 13, 2026