Kubernetes 1.37, released on August 26, 2026, expands the platform’s Dynamic Resource Allocation (DRA) capabilities. DRA itself has been stable since Kubernetes 1.35, but the 1.37 release adds and changes several related feature gates for workload claims, device compatibility, derived attributes, node operations, and partitionable-device types.
For platform teams, the important distinction is maturity. Core DRA allocation is stable, while several of the newer capabilities remain alpha or beta, and some are disabled by default. Adoption therefore depends on the installed DRA driver, administrator-managed feature gates, workload definitions, and carefully scoped permissions.
What changed in Kubernetes 1.37
The Kubernetes 1.37 release series is actively supported. Its DRA-related feature surface includes the following maturity levels and defaults, according to the official feature-gate reference:
- DRAWorkloadResourceClaims is beta and disabled by default.
- DRADeviceCompatibilityGroups is alpha and disabled by default.
- DRADerivedAttributes is alpha and disabled by default.
- DRAOptionalNodeOperations is alpha and disabled by default.
- DRAPartitionableDevicesType is alpha and disabled by default.
- DRAConsumableCapacity is beta and enabled by default.
- DRADeviceBindingConditions is beta and enabled by default.
Partitionable devices are already beta as of Kubernetes 1.36 and remain enabled by default in 1.37. The DRAPartitionableDevices gate controls the capability in the kube-apiserver and kube-scheduler.
Device compatibility groups build on partitionable-device support. The feature can help scheduling account for incompatible combinations of partitioned devices instead of leaving all compatibility decisions to node-side preparation. It requires both DRADeviceCompatibilityGroups and DRAPartitionableDevices to be enabled in the kube-apiserver and kube-scheduler.
Stable allocation, expanding workload options
The stable foundation is DRA device allocation. Kubernetes administrators configure DRA, attach devices, and install compatible DRA drivers. Workloads then request those devices through ResourceClaim or ResourceClaimTemplate objects.
A ResourceClaim can be referenced by multiple Pods, which supports workloads that need to share an allocated device or resource claim. A ResourceClaimTemplate instead allows separately generated claims, including for workload patterns such as parallel Jobs. In a Pod specification, claims are referenced through the Pod’s resourceClaims field and the relevant container resource claim fields.
The beta DRAWorkloadResourceClaims capability is disabled by default in 1.37, so teams should not assume that its behavior is available merely because the cluster runs the new release. Administrators need to verify the gate state and confirm that the installed driver supports the required workload behavior before changing production manifests.
Feature gates require deliberate administration
Feature gates are configured with the --feature-gates flag as comma-separated key-value pairs on applicable Kubernetes components. However, the component mapping is not identical for every DRA-related feature.
The official DRA feature documentation specifically identifies the kube-apiserver and kube-scheduler for partitionable devices and device compatibility groups. For compatibility groups, both related gates must be configured on those components. The available documentation does not establish a complete component-by-component mapping for every DRA gate in 1.37, so operators should verify the relevant feature’s documentation rather than applying a blanket configuration to all control-plane services.
This distinction matters operationally. Enabling an alpha gate can change scheduling or allocation behavior while the feature is still experimental. A gate should be enabled only when the cluster version, DRA driver, and workload design have been checked together.
Security and authorization boundaries
DRA drivers and controllers update ResourceClaim status, so those components need explicit, least-privilege authorization. Kubernetes’ DRA hardening guidance describes synthetic subresources for separating these operations.
The resourceclaims/binding subresource controls updates to allocation and reservation status, including status.allocation and status.reservedFor. The resourceclaims/driver subresource controls device status updates. Node-local drivers can use associated-node verbs, while multi-node or control-plane controllers may require arbitrary-node verbs.
Arbitrary-node permissions have broader scope and should be granted only when necessary. The guidance establishes authorization boundaries for DRA status updates; it is not a general security guarantee for every driver or deployment. Teams should review which component needs which subresource and whether its operations are limited to an associated node.
What developers and administrators should do
- Separate stable usage from preview usage. Treat standard DRA allocation as the stable baseline. Review the maturity and default state of each alpha or beta capability before using it.
- Confirm the cluster prerequisites. DRA requires administrator setup, attached devices, and compatible DRA drivers. The documented workload task requires a Kubernetes server version of 1.34 or later.
- Choose the right claim model. Use a shared
ResourceClaimwhen multiple Pods should reference the same claim. Use aResourceClaimTemplatewhen workloads need separately generated claims. - Validate driver support. A Kubernetes feature gate does not by itself ensure that the installed DRA driver supports the capability required by a workload.
- Check gates on the correct components. In particular, verify kube-apiserver and kube-scheduler settings for partitionable devices and device compatibility groups.
- Review RBAC before deployment. Grant only the DRA status-update permissions required by each driver or controller, using the appropriate synthetic subresource and node-aware scope.
Teams evaluating a move from device plugins to DRA should avoid treating Kubernetes 1.37 as a complete migration guide. The supplied Kubernetes documentation establishes the DRA workload model and administration requirements, but it does not provide a complete device-plugin-to-DRA migration procedure.
Availability and remaining limits
Kubernetes 1.37 is available, and the 1.37 release series is actively supported. DRA allocation is stable, while partitionable devices are beta and enabled by default. Workload claims are beta but disabled by default, and device compatibility groups, derived attributes, optional node operations, and partitionable-device types are alpha and disabled by default according to the v1.37 feature-gate reference.
The exact operational effect of each advanced feature depends on the DRA driver and the Kubernetes components supporting its gate. The official documentation also does not establish complete component mappings for every DRA-related feature. Operators should therefore validate each capability in a controlled environment before relying on it in production workloads.
