Hardening Guide - Dynamic Resource Allocation

فهرست فصل‌ها دانلود آرشیوناشر

Kubernetes Docs · نسخهٔ f284a9e7c87f · CC BY 4.0 برای مستندات و مجوزهای همراه

<!-- overview -->

Dynamic Resource Allocation (DRA) adds powerful scheduling and device management capabilities. Because DRA components update ResourceClaim status, cluster administrators should configure authorization for those updates with explicit, least-privilege RBAC.

{{< feature-state feature_gate_name="DRAResourceClaimGranularStatusAuthorization" >}}

Starting in Kubernetes v1.36, DRA status updates use synthetic subresources and, in some cases, specialized node-aware verbs.

<!-- body -->

Harden DRA status update permissions

For DRA status updates,In addition to granting update permissions on the resourceclaims/status subresource, cluster administrators must grant permissions on specific "synthetic" subresources based on the exact fields a component needs to modify. This enforces the principle of least privilege between the scheduler, custom controllers, and DRA drivers.

The DRA authorization checks are divided into two synthetic subresources:

  • **resourceclaims/binding**
  • Required to modify status.allocation and status.reservedFor.
  • Typically granted to the kube-scheduler and custom allocation controllers.
  • Uses standard update and patch verbs.
  • **resourceclaims/driver**
  • Required to modify status.devices.
  • This check is performed per-driver to drivers from tampering with devices on different

nodes and/or from other drivers.

  • Uses node-aware verbs for stricter scope.

Node-aware DRA verbs

When authorizing updates to resourceclaims/driver, use the appropriate specialized verb prefix:

  • **associated-node:<verb>** (for example, associated-node:update)
  • For node-local drivers.
  • The API server verifies node association for the requesting driver.
  • **arbitrary-node:<verb>** (for example, arbitrary-node:patch)
  • For control-plane or multi-node controllers that may update claims from

any node.

Example RBAC patterns

Scheduler and allocation controller permissions

yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: dra-binding-updater
rules:
  - apiGroups: ["resource.k8s.io"]
    resources: ["resourceclaims/status"]
    verbs: ["get", "patch", "update"]
  - apiGroups: ["resource.k8s.io"]
    resources: ["resourceclaims/binding"]
    verbs: ["patch", "update"]

Node-local DRA driver permissions

yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: dra-node-driver-status-updater
rules:
  - apiGroups: ["resource.k8s.io"]
    resources: ["resourceclaims/status"]
    verbs: ["get", "patch", "update"]
  - apiGroups: ["resource.k8s.io"]
    resources: ["resourceclaims/driver"]
    verbs: ["associated-node:patch", "associated-node:update"]
    resourceNames: ["dra.example.com"]

Multi-node status controller permissions

yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: dra-multinode-status-updater
rules:
  - apiGroups: ["resource.k8s.io"]
    resources: ["resourceclaims/status"]
    verbs: ["get", "patch", "update"]
  - apiGroups: ["resource.k8s.io"]
    resources: ["resourceclaims/driver"]
    verbs: ["arbitrary-node:patch", "arbitrary-node:update"]
    resourceNames: ["dra.example.com"]

Related cluster administrator task

To apply these patterns in a running cluster, see Harden Dynamic Resource Allocation in Your Cluster.

{{% heading "whatsnext" %}}

  • Authorization
  • Set Up DRA in a Cluster
  • Dynamic Resource Allocation