DRA Preemption

Dynamic Resource Allocation (DRA) was promoted to stable in v1.35, but one major capability—preemption of pods utilizing DRA resources—is currently not supported. Under the current state, lower-priority workloads that happen to be scheduled early can occupy scarce devices indefinitely, leaving higher-priority workloads pending. By introducing native preemption support in DRA, we equip the scheduler with the necessary capabilities to ensure that the most critical workloads gain access to scarce hardware resources. Simulating the removal of victims is not sufficient on its own. The devices held by preempted pods are reclaimed asynchronously by the resourceclaim controller, so there is a window during which a preemptor has been promised capacity that it does not yet hold. This KEP therefore also specifies how the scheduler records and holds that capacity, so that a preemptor is not made to preempt repeatedly and its promised resources are not taken by other pods while being reclaimed. Most of this is contained in the dynamicresources plugin. The one exception is a new optional scheduler framework extension point, NominationExtensions, which notifies plugins when a preemption candidate is selected or cleared and lets a plugin report whether resources freed by an earlier preemption are still being reclaimed.