Kubernetes 1.34: Dynamic Resource Allocation (DRA) reaches general availability

2025-08-27 · Publisher : Kubernetes · Domain : Cloud & DevOps

On 27 August 2025, Kubernetes 1.34 ships 58 enhancements (23 stable); the core of DRA, which manages GPUs and other accelerators, becomes stable and enabled by default, with a no-breaking-change promise.

### Background On 27 August 2025, Kubernetes releases version **1.34 "Of Wind & Will"**: **58 enhancements**, including 23 stable, 22 beta and 13 alpha. The headline: the core of **DRA** (Dynamic Resource Allocation), the framework that lets Kubernetes manage specialized hardware such as GPUs or FPGAs, reaches general availability. ### Key facts - `resource.k8s.io` group APIs (ResourceSlice, DeviceClass, ResourceClaim) move to stable v1; DRA is enabled by default, as are its beta features. - According to the DRA team, new features will keep coming in coming releases but without breaking changes: driver authors can adopt DRA "with confidence". - In beta: admin-access labeling against privilege escalation, and a prioritized device list (one high-end GPU, otherwise two mid-range ones). - The kubelet PodResources API reports resources allocated through DRA, which node monitoring agents can use. - Other beta graduations: service account tokens for image-pull authentication and pod-level resource requests. ### Why it matters GPUs are the scarcest and most expensive resource in AI clusters. DRA replaces the old device-plugin model with a scheme where each workload describes the properties it needs and the scheduler picks the device. For an IT services firm, this improves GPU utilization and makes AI platforms more portable between providers. Vendors' DRA drivers still need qualifying one by one. ### Takeaway Stable DRA makes Kubernetes a first-rank GPU scheduler, with an API that will no longer break.

Tags : Kubernetes, IA, Cloud natif, Puces, Open Source