Files
runner-container-hooks/docs/adrs/0135-rwx-volume-strategy.md
T
Nikola JokicandSisyphus 51a0a891f6 docs(adr,k8s): clarify RWX enables free scheduling, RWO requires affinity
- RWX volumes allow job pods to be scheduled on any cluster node
- RWO volumes require affinity to pin job pods to runner's node
- Remove ACTIONS_RUNNER_USE_KUBE_SCHEDULER from RWX migration steps
- Emphasize resource utilization benefits of RWX free scheduling

Co-authored-by: Sisyphus <[email protected]>
2026-04-23 00:09:05 +02:00

4.8 KiB

ADR 0135: RWX volume strategy and RWO affinity fallback

Date: 22 April 2026

Status: Accepted

Context

The Kubernetes hook implementation for GitHub Actions runners requires access to the runner's working directory (_work) within the dynamically created job pods. This shared access is typically managed via Persistent Volume Claims (PVCs).

The choice of storage strategy significantly impacts pod scheduling. While ReadWriteOnce (RWO) volumes require job pods to be co-located on the same node as the runner pod, ReadWriteMany (RWX) volumes allow job pods to be scheduled freely across the cluster.

Depending on the storage provider and cluster configuration, operators may choose between ReadWriteMany (RWX) or ReadWriteOnce (RWO) access modes. RWX is preferred because it allows the Kubernetes scheduler to place job pods on any available node, improving resource utilization and cluster flexibility. RWO restricts volume access to a single node at a time, requiring all pods using the volume to be pinned to that specific node.

Decision

We have decided to establish ReadWriteMany (RWX) as the preferred storage strategy for the Kubernetes hook. RWX provides superior operational flexibility by enabling free scheduling of job pods across the cluster, as the shared volume is accessible from any node. This decoupling of job pods from the runner's node allows for better resource distribution and reduces the risk of node-level resource exhaustion.

For environments where RWX is unavailable or undesirable, we support a ReadWriteOnce (RWO) fallback strategy. This fallback is implemented using node affinity to ensure that job pods are scheduled onto the same node as the runner pod that holds the RWO volume.

Operational Guidance

  1. Preferred Model (RWX): Operators should configure the runner with a PVC supporting ReadWriteMany.
  2. Fallback Model (RWO): If using ReadWriteOnce, operators must enable the Kubernetes scheduler integration by setting ACTIONS_RUNNER_USE_KUBE_SCHEDULER=true.
  3. Node Selection: When scheduler integration is enabled, the hook applies a requiredDuringSchedulingIgnoredDuringExecution node affinity targeting the runner's current node (kubernetes.io/hostname).
  4. Implementation Details:
    • The hook determines the node name via getCurrentNodeName() and applies affinity in packages/k8s/src/k8s/index.ts (lines 101, 165).
    • The scheduler behavior is toggled by the ACTIONS_RUNNER_USE_KUBE_SCHEDULER environment variable, as defined in packages/k8s/src/k8s/utils.ts (line 16).
    • The PVC claim name defaults to ${ACTIONS_RUNNER_POD_NAME}-work unless overridden by ACTIONS_RUNNER_CLAIM_NAME (packages/k8s/src/hooks/constants.ts, lines 27-33).

Non-Recommendations

We explicitly do not recommend the use of spec.nodeName for operator-driven scheduling. While the hook uses nodeName as a legacy fallback when ACTIONS_RUNNER_USE_KUBE_SCHEDULER is not set to true (packages/k8s/src/k8s/index.ts, lines 103, 167), this bypasses the Kubernetes scheduler and can lead to scheduling failures or resource imbalances. Operators should always prefer the affinity-based approach for RWO volumes.

Alternatives

  • nodeName Bypass: Directly setting nodeName bypasses the scheduler entirely. This was rejected as a recommendation because it prevents the scheduler from accounting for taints, tolerations, and resource pressure.
  • Local Volumes: Using local volumes tied to specific nodes. This is a subset of the RWO fallback and is supported via the affinity mechanism.

Consequences

  • Flexibility: RWX users benefit from the ability to schedule job pods on any node in the cluster, maximizing resource utilization.
  • Node Coupling: RWO users remain coupled to the node where the runner pod is running. The hook ensures job pods are scheduled on the same node via affinity to maintain workspace integrity.
  • Configuration: Operators must be aware of the ACTIONS_RUNNER_USE_KUBE_SCHEDULER toggle when using RWO. This toggle controls whether the hook uses nodeName (bypassing the scheduler) or node affinity (using the scheduler) to pin the pod to the runner's node. RWX configurations do not require this toggle for basic operation.

Migration Guidance

Operators migrating from an RWO setup that relied on default nodeName behavior to a more robust affinity-based setup should:

  1. Ensure the runner pod has the ACTIONS_RUNNER_USE_KUBE_SCHEDULER environment variable set to true.
  2. Verify that the runner's ServiceAccount has the necessary permissions to list pods (to determine its own node).

Non-Goals

  • This ADR does not recommend nodeName as a primary or secondary configuration path for operators.
  • This ADR does not dictate specific storage providers (e.g., EBS vs. EFS vs. Azure Files), but rather the access mode strategy.