Runner: place it by node label and give DinD capped local disk
Demo / hello (push) Successful in 27s

- Schedule the runner only on a node labeled eom.dev/ci-runner=true,
  and fail the deploy with a clear message if no node has the label.
- Put DinD's /var/lib/docker on an emptyDir with a 50Gi sizeLimit, so
  Docker data uses local disk (overlay2 doesn't work on the NFS storage
  classes) and has a known cap.
- Request 20Gi and limit 55Gi of ephemeral storage on DinD, which also
  keeps the scheduler off small-disk nodes.

All sizes are role defaults. Which node gets the label is Eric's call.

Refs DevOps/software-infrastructure#36

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-11-09 20:03:28 -05:00
co-authored by Claude Opus 5.5
parent ffd782cfa6
commit e83262d862
2 changed files with 49 additions and 0 deletions
+18
View File
@@ -40,3 +40,21 @@ gitea_actions_enabled: true
gitea_actions_chart_version: 0.1.2
gitea_actions_runner_replicas: 1
gitea_actions_runner_token_secret: gitea-actions-runner-token
# Where the runner runs and how much disk Docker-in-Docker may use.
# The runner only schedules on a node carrying this label. Choose a node with
# real local disk (the cluster's NFS storage classes don't suit overlay2) and
# label it once:
# kubectl label node <node> eom.dev/ci-runner=true
# The role checks for the label before deploying. Labels live on the node, so
# re-label it if the node is rebuilt.
gitea_actions_runner_node_selector:
eom.dev/ci-runner: "true"
# DinD's /var/lib/docker goes on an emptyDir on that node's local disk, capped
# here. The kubelet evicts the pod if the cap is exceeded, so S5's cleanup has
# to keep usage well below it. One runner image is about 5 GB.
gitea_actions_docker_data_size: 50Gi
# Ephemeral storage for the DinD container, which holds the emptyDir above.
# The request also keeps the scheduler off nodes with small disks.
gitea_actions_dind_ephemeral_request: 20Gi
gitea_actions_dind_ephemeral_limit: 55Gi
+31
View File
@@ -45,6 +45,22 @@
token: "{{ runner_token.stdout | trim }}"
no_log: true
- name: find the node labeled for the runner
kubernetes.core.k8s_info:
kind: Node
label_selectors: "{{ gitea_actions_runner_node_selector.items() | map('join', '=') | list }}"
register: runner_nodes
- name: require a node labeled for the runner
ansible.builtin.assert:
that: runner_nodes.resources | length > 0
fail_msg: >-
No node matches {{ gitea_actions_runner_node_selector }}. Label a node
with enough local disk for Docker-in-Docker, for example
`kubectl label node <node> eom.dev/ci-runner=true`, then rerun with
--tags actions.
quiet: true
- name: deploy gitea actions runner
kubernetes.core.helm:
name: gitea-actions
@@ -61,6 +77,21 @@
replicas: "{{ gitea_actions_runner_replicas }}"
persistence:
size: 1Gi
# Chart 0.1.2 renders nodeSelector correctly for one key only.
nodeSelector: "{{ gitea_actions_runner_node_selector }}"
extraVolumes:
- name: docker-data
emptyDir:
sizeLimit: "{{ gitea_actions_docker_data_size }}"
dind:
extraVolumeMounts:
- name: docker-data
mountPath: /var/lib/docker
resources:
requests:
ephemeral-storage: "{{ gitea_actions_dind_ephemeral_request }}"
limits:
ephemeral-storage: "{{ gitea_actions_dind_ephemeral_limit }}"
runner:
config: |
log: