The Kubernetes Device Management Working Group (WG Device Management) is spotlighted in an interview with its three co-chairs from NVIDIA, Intel, and Google. The group's primary deliverable, Dynamic Resource Allocation (DRA), recently graduated to GA in Kubernetes 1.34. DRA replaces the legacy Device Plugin API — which treated devices as opaque integers — with a structured, declarative framework covering four stages: modeling (ResourceSlice API), requesting (ResourceClaim API), scheduling, and actuation. This enables workloads to specify fine-grained hardware requirements like GPU memory, interconnect topology, and partitioning. The working group spans five SIGs (node, scheduling, autoscaling, network, architecture) and is now focused on extending DRA expressiveness, day-two operations like health monitoring, multi-node topology-aware scheduling, and consumable capacity sharing models. Key challenges include NP-hard scheduling complexity and modeling diverse hardware. NVIDIA recently donated its DRA GPU driver to the Kubernetes project, and the community is growing rapidly.
