Kubernetes v1.36 introduces an architectural evolution in workload-aware scheduling by decoupling the Workload API from the PodGroup API. This separation enables the Workload API to function as a static template while the PodGroup API manages runtime state, facilitating atomic scheduling and improved scalability.
Key points
Decoupling the Workload API from the PodGroup API allows the scheduler to operate directly on runtime objects, improving performance.
The new PodGroup API supports per-replica sharding of status updates, enhancing scalability for large batch workloads.
Workload-aware scheduling now includes topology-aware scheduling, workload-aware preemption, and integrated support for Dynamic Resource Allocation.
The Job controller integration with the new API enables native support for atomic, gang-scheduled workloads.
