Dynamic Resource Allocation (DRA) continues to reshape how cluster operators handle specialized hardware in Kubernetes. With the v1.36 release, DRA moves further into maturity, introducing stable and beta features that improve scheduling flexibility, hardware utilization, and migration paths from legacy resource models. New capabilities allow you to define fallback preferences, partition powerful accelerators, and even taint faulty devices. The ecosystem of supported drivers also expands beyond GPUs to include networking and other hardware types. Whether you're managing large fleets of AI accelerators or looking for better failure handling, this update delivers tangible improvements. Below we answer the most pressing questions about what's new in DRA with Kubernetes v1.36.

1. What is Dynamic Resource Allocation (DRA) and why is it central to Kubernetes v1.36?

Dynamic Resource Allocation (DRA) is a Kubernetes framework that enables fine-grained control over hardware resources such as GPUs, FPGAs, and network devices. Instead of relying on static resource requests, DRA uses ResourceClaims to dynamically allocate specific hardware to pods based on policy. In v1.36, DRA graduates several key features to stable and beta, making it the de facto standard for managing specialized hardware. The release also extends DRA to support native resources like memory and CPU, and introduces support for ResourceClaims in PodGroups. This maturation is critical because it allows platform administrators to handle heterogeneous hardware environments more flexibly, improves scheduling reliability through features like device binding conditions, and eases transitions from legacy resource models. Ultimately, DRA empowers clusters to maximize utilization of expensive accelerators while maintaining strict isolation and policy controls.

2. How does the new Prioritized List feature improve scheduling flexibility?

The Prioritized List feature, now stable in v1.36, addresses hardware heterogeneity by allowing you to define an ordered list of preferred device types. For instance, you can specify "Give me an NVIDIA H100 if available, otherwise fall back to an A100." The scheduler evaluates these preferences in order, trying to assign the first matching device; if none are free, it moves down the list. This drastically improves scheduling flexibility because pods no longer get stuck waiting for a specific model when alternatives exist. Cluster utilization increases as hard-to-find high-end accelerators are used only when necessary, while more common devices handle the rest. Administrators can define fallback chains that match their hardware inventory, reducing fragmentation and enabling better resource sharing across diverse workloads.

3. What is Extended Resource support and how does it facilitate migration to DRA?

The Extended Resource support feature, now in beta, bridges the gap between the traditional extended resources model and DRA. It allows pods to request resources via the familiar extended resources mechanism while DRA handles the actual allocation under the hood. This enables a gradual, non-disruptive migration: cluster operators can enable DRA on the infrastructure side without forcing application developers to immediately adopt the ResourceClaim API. Developers continue using resources.requests for extended resources as before, while the cluster translates these requests into DRA allocations. This backward compatibility reduces friction and risk, allowing teams to move at their own pace while still benefiting from DRA's advanced scheduling and device management capabilities.

4. How do Partitionable Devices optimize hardware utilization?

The Partitionable Devices feature (beta) allows DRA to dynamically carve powerful hardware accelerators, such as GPUs with Multi-Instance GPU (MIG) support, into smaller logical units. Instead of allocating an entire high-end GPU to a single pod, administrators can define a device as partitionable and let the scheduler assign fractions (e.g., one-third of the GPU) to multiple pods. This dramatically improves utilization in environments where workloads don't need the full device capacity. The feature works by exposing device slices as separate resources, which can be claimed independently. It also maintains isolation between partitions, ensuring that workloads do not interfere with each other. For expensive hardware, this means more pods can share the same physical device, reducing costs and enabling better density in AI/ML clusters.

5. What are Device Taints and how do they improve hardware management?

Inspired by node taints, Device Taints (beta) enable you to apply taints directly to individual DRA devices. This allows cluster administrators to mark specific hardware as faulty, reserved, or dedicated to certain workloads. For example, a GPU with intermittent errors can be tainted so only pods with matching tolerations can claim it—ideal for testing or non-critical jobs. Similarly, you can reserve high-end accelerators for priority workloads by tainting all others and only allowing tolerations on those pods. This fine-grained control prevents accidental allocation of degraded or restricted hardware, improves reliability, and supports multi-tenant or team-based resource allocation policies. Device taints integrate seamlessly with existing Kubernetes taint/toleration concepts, making them intuitive for operators.

6. How do Device Binding Conditions enhance scheduling reliability?

Device Binding Conditions (beta) add a new layer of scheduling reliability by allowing devices to report their operational status before being bound to a pod. When a pod requests a device, the scheduler checks binding conditions that may include health checks, driver availability, or prerequisite states. If the device does not meet the conditions, the scheduler can delay binding or choose an alternative. This prevents scenarios where a pod is scheduled onto a device that later fails to initialize, reducing wasted scheduling cycles and improving overall cluster stability. For operators, this means fewer retries and better predictability, especially in environments with complex hardware dependencies. The feature also supports custom conditions via vendor-specific logic, making it extensible.

7. What are the implications of DRA maturing for GPU and hardware management?

With DRA's graduation of core features in v1.36, managing GPUs and other specialized hardware becomes more robust and flexible. The combination of prioritized lists, partitionable devices, device taints, and binding conditions gives operators granular control over allocation, failure handling, and utilization. The extended resource support eases adoption without disrupting existing workflows. Furthermore, the driver ecosystem now includes networking hardware and other device types, moving towards a hardware-agnostic infrastructure. For AI/ML workloads, this means better performance isolation, reduced idle costs, and the ability to share expensive accelerators safely. Overall, DRA's maturation signals that Kubernetes is ready for production-grade heterogeneous hardware management, enabling more efficient cluster operations at scale.