Kubernetes v1.36 DRA Overhaul: Prioritized Lists, Device Taints, and Partitionable GPUs Go Stable/Beta

Breaking: Kubernetes v1.36 Delivers Major DRA Upgrades—Prioritized Lists Now Stable, Extended Resources in Beta

San Francisco, CA — The Kubernetes community today released v1.36, bringing landmark advancements to Dynamic Resource Allocation (DRA), including the graduation of prioritized device lists to stable and the introduction of beta features for extended resources, partitionable devices, device taints, and binding conditions. The release marks a pivotal step toward hardware-agnostic, flexible resource management for cloud-native workloads.

Prioritized Lists Go Stable

The most anticipated feature, Prioritized lists, now stable, lets administrators define fallback preferences for hardware accelerators. For example, a pod can request an NVIDIA H100, but if none are available, automatically fall back to an A100. This flexibility dramatically improves cluster utilization and scheduling success.

“Prioritized lists solve a real pain point for teams managing heterogeneous GPU fleets,” said Priya Sharma, senior Kubernetes contributor at CloudNative Labs. “Instead of rigid requests, you now have fine-grained control over device selection without custom schedulers.”

Extended Resource Support (Beta)

To ease migration from legacy systems, DRA now supports Extended resources as a beta feature. This allows pods to request resources via traditional extended resource names while gradually adopting the ResourceClaim API. It bridges the gap between old and new allocation methods.

Partitionable Devices (Beta)

For multi-instance GPU (MIG) scenarios, Partitionable devices enable dynamic carving of physical hardware into logical slices. A single GPU can now be safely shared across multiple pods, reducing waste and cost.

Device Taints and Tolerations (Beta)

Inspired by node taints, Device taints allow admins to mark faulty or reserved hardware. Only pods with matching tolerations can claim tainted devices—useful for isolating experimental workloads or quarantining defective accelerators.

Device Binding Conditions (Beta)

To improve scheduling reliability, Device binding conditions ensure that resource claims are only bound when the hardware is fully ready. This prevents race conditions where a pod starts before the accelerator is initialized.

Background

Dynamic Resource Allocation (DRA) was initially introduced in Kubernetes v1.26 as an alpha feature to manage specialized hardware like GPUs, FPGAs, and network accelerators. Over successive releases, it has matured from a niche subsystem to a core scheduling component. The v1.36 release builds on feedback from large-scale deployments at tech giants and research institutions, focusing on operational simplicity and cluster efficiency.

What This Means

For platform engineers, these changes reduce the need for custom device plugins and manual resource orchestration. The stable prioritized list alone can increase GPU utilization by up to 30% in heterogeneous clusters, according to community benchmarks. For application developers, extended resource support means a smoother transition path—they can keep using familiar extended resource requests while the backend shifts to DRA. Partitionable devices promise better ROI on expensive accelerators, and device taints offer fine-grained access control.

“This release solidifies DRA as the go-to solution for next-generation resource management in Kubernetes,” added Michael Torres, senior engineer at a major cloud provider. (Read more background)

How to Upgrade

Cluster operators can enable DRA features via feature gates in Kubernetes v1.36. Refer to the official DRA documentation for configuration details.