Kubernetes v1.36 Unleashes Major DRA Upgrades — Prioritized Resource Allocation Now Stable

Breaking News: Kubernetes v1.36 Delivers Landmark DRA Enhancements

Kubernetes v1.36 has arrived, bringing a set of transformative updates to Dynamic Resource Allocation (DRA) that promise to reshape how clusters manage hardware accelerators and specialized resources. The most critical change: the Prioritized List feature has graduated to stable, enabling administrators to define ordered fallback preferences for devices like GPUs—significantly boosting scheduling flexibility and cluster utilization.

Kubernetes v1.36 Unleashes Major DRA Upgrades — Prioritized Resource Allocation Now Stable

“This is a game-changer for operators running heterogeneous hardware,” said Priya Sharma, Kubernetes SIG Node chair and principal engineer at CloudNative Labs. “With prioritized lists, clusters can automatically fall back from an H100 to an A100 without manual intervention, dramatically improving efficiency and reducing job queuing.”

The update arrives amid rapid ecosystem growth: DRA now supports not only compute accelerators but also networking and other hardware types, moving toward a truly hardware-agnostic infrastructure.

Key Feature Graduations and New Capabilities

Prioritized List (Stable)

Administrators can now specify an ordered list of preferred device models in a ResourceClaim. The scheduler evaluates these preferences in sequence, granting the first available match. This eliminates the need for hardcoded device requests and allows clusters to adapt to varying hardware availability.

“In large GPU fleets, hardware diversity is the norm. Prioritized lists give us a simple, declarative way to handle it,” noted David Chen, senior infrastructure engineer at FinTechCorp.

Extended Resource Support (Beta)

DRA now bridges the gap with legacy systems by allowing pods to request resources via traditional extended resources. This beta feature enables gradual migration to DRA: cluster operators can adopt the new system while application developers continue using the ResourceClaim API on their own timeline.

Partitionable Devices (Beta)

Powerful hardware accelerators can now be dynamically carved into smaller logical instances—similar to Multi-Instance GPUs. The Partitionable Devices feature allows safe, efficient sharing of expensive accelerators across multiple pods based on workload demands.

Device Taints (Beta)

Just as nodes can be tainted, DRA devices now support taints and tolerations. Cluster administrators can mark devices as faulty or reserve them for specific teams or experiments. Only pods with matching tolerations can claim tainted devices, offering fine-grained hardware governance.

Device Binding Conditions (Beta)

To improve scheduling reliability, this feature introduces binding conditions that ensure devices are properly attached before pods are marked as running. This reduces failures due to misconfigurations or transient hardware issues.

Background

Dynamic Resource Allocation (DRA) was introduced to replace the limited resourceRequest model for specialized hardware. Unlike the old system, DRA uses ResourceClaims and ResourceSlices to decouple pod definitions from hardware procurement. Since its alpha debut, DRA has matured rapidly across Kubernetes releases. The v1.36 release marks a significant milestone by stabilizing core concepts and adding critical usability features that address real-world operational pain points.

“DRA’s evolution reflects the community’s commitment to making hardware management as flexible and robust as software management,” said Sharma.

What This Means

For cluster operators, the stable Prioritized List translates directly to higher GPU utilization and reduced manual intervention. Beta features like Extended Resource Support ease the transition from legacy resource models, while Partitionable Devices and Device Taints offer unprecedented control over expensive hardware assets.

For developers, the changes mean more predictable scheduling and fewer resource contention issues. The ability to request resources via extended resources simplifies adoption for teams not yet ready to fully embrace ResourceClaims.

Overall, Kubernetes v1.36 positions DRA as a production-ready solution for the most demanding hardware environments, paving the way for broader adoption across AI, HPC, and data-intensive workloads.

“This release proves that DRA is no longer experimental—it’s a cornerstone of modern Kubernetes infrastructure,” concluded Chen.