Support Node affinity for pause pods with Kubernetes executor

Summary

I wanted to try pause pods to speed up jobs startup. However the pause pods don't spawn on the right nodes. They seem to ignore the runners.kubernetes.affinity.node_affinity configuration.

Steps to reproduce

1. Configure a Kubernetes runner with runners.kubernetes.affinity.node_affinity
2. Enable pause pods

Actual behavior

The pause pods deployment ignores the node affinity

Expected behavior

The pause pods should deploy on the same nodes as the jobs will

Relevant logs and/or screenshots

In the following manifest, we can see the configured node toleration is here, but not the node affinity :

runner-pause deployment YAML

apiVersion: apps/v1
kind: Deployment
metadata:
  annotations:
    deployment.kubernetes.io/revision: '1'
    runner.gitlab.com/pause-heartbeat: '2026-07-21T09:05:43Z'
  creationTimestamp: '2026-07-21T08:52:53Z'
  generation: 12
  labels:
    manager.runner.gitlab.com/id-short: kUpb21O-A
    manager.runner.gitlab.com/system-id: r_oIQUNuA5lMFH
    runner.gitlab.com/pause-pod: 'true'
  name: runner-pause-kupb21o-a-roiqunua5lmfh
  namespace: gitlab
  resourceVersion: '1784624743122959000'
  uid: 827b7e7d-5124-45de-a558-d558cd2b6c4c
spec:
  progressDeadlineSeconds: 600
  replicas: 5
  revisionHistoryLimit: 10
  selector:
    matchLabels:
      manager.runner.gitlab.com/id-short: kUpb21O-A
      manager.runner.gitlab.com/system-id: r_oIQUNuA5lMFH
      runner.gitlab.com/pause-pod: 'true'
  strategy:
    rollingUpdate:
      maxSurge: 25%
      maxUnavailable: 25%
    type: RollingUpdate
  template:
    metadata:
      labels:
        manager.runner.gitlab.com/id-short: kUpb21O-A
        manager.runner.gitlab.com/system-id: r_oIQUNuA5lMFH
        runner.gitlab.com/pause-pod: 'true'
    spec:
      containers:
      - image: registry.k8s.io/pause:3.10
        imagePullPolicy: IfNotPresent
        name: pause
        resources:
          requests:
            cpu: 500m
            memory: 1Gi
        terminationMessagePath: /dev/termination-log
        terminationMessagePolicy: File
      dnsPolicy: ClusterFirst
      priorityClassName: pause-pods-priority-class
      restartPolicy: Always
      schedulerName: default-scheduler
      securityContext: {}
      serviceAccount: default
      serviceAccountName: default
      terminationGracePeriodSeconds: 0
      tolerations:
      - effect: NoSchedule
        key: preemptible
        operator: Equal
        value: 'true'

Environment description

I am using Kubernetes executor on GCP with a self-hosted GitLab instance. We use node affinity in order for jobs to prefer less expensive spot nodes if available.

config.toml contents

[runners.kubernetes.autoscaler]
  max_pause_pods = 10
  pause_pod_priority_class_name = "pause-pods-priority-class"
  [[runners.kubernetes.autoscaler.policy]]
    idle_count = 0
    periods = ["* * * * *"]
  [[runners.kubernetes.autoscaler.policy]]
    idle_count = 5
    periods = ["* 7-20 * * mon-fri"]

[runners.kubernetes.node_tolerations]
"preemptible=true" = "NoSchedule"

# Set preference for spot VM but will fallback to non-spot if no spot VMs are available
[runners.kubernetes.affinity.node_affinity]
  [[runners.kubernetes.affinity.node_affinity.preferred_during_scheduling_ignored_during_execution]]
    weight = 100
    [runners.kubernetes.affinity.node_affinity.preferred_during_scheduling_ignored_during_execution.preference]
      [[runners.kubernetes.affinity.node_affinity.preferred_during_scheduling_ignored_during_execution.preference.match_expressions]]
        key = "cloud.google.com/gke-spot"
        operator = "In"
        values = ["true"]
  [runners.kubernetes.affinity.node_affinity.required_during_scheduling_ignored_during_execution]
    [[runners.kubernetes.affinity.node_affinity.required_during_scheduling_ignored_during_execution.node_selector_terms]]
      [[runners.kubernetes.affinity.node_affinity.required_during_scheduling_ignored_during_execution.node_selector_terms.match_expressions]]
        key = "ci"
        operator = "In"
        values = ["true"]

Used GitLab Runner version

Docker image : gitlab/gitlab-runner:v19.2.0

Possible fixes

I didn't find anything related to node affinity in the pause_pod_manager

I guess I could mitigate the issue by using node selectors / tolerations, but then having a preference for spot nodes is impossible