The runner has no CPU or memory requests or limits and runs two jobs at once, on nodes that mostly have 1 CPU. It uses default job and shutdown timeouts.
Work
Add CPU and memory requests and limits for the runner and DinD containers.
Set capacity to what the chosen node can sustain (likely 1 until a build host exists).
Set runner.timeout, shutdown_timeout, fetch_timeout, and a matching terminationGracePeriodSeconds so jobs can finish or report during a restart.
Set container.force_pull: false and pin the three runner image labels instead of latest where practical.
Acceptance
Values live in role defaults and the rendered config is shown in a comment.
A forced pod restart produces a clear job failure rather than a silent one.
Depends on: S2. Label ready-to-deploy when committed.
The runner has no CPU or memory requests or limits and runs two jobs at once, on nodes that mostly have 1 CPU. It uses default job and shutdown timeouts.
**Work**
- Add CPU and memory requests and limits for the runner and DinD containers.
- Set `capacity` to what the chosen node can sustain (likely 1 until a build host exists).
- Set `runner.timeout`, `shutdown_timeout`, `fetch_timeout`, and a matching `terminationGracePeriodSeconds` so jobs can finish or report during a restart.
- Set `container.force_pull: false` and pin the three runner image labels instead of `latest` where practical.
**Acceptance**
- Values live in role defaults and the rendered config is shown in a comment.
- A forced pod restart produces a clear job failure rather than a silent one.
Depends on: S2. Label `ready-to-deploy` when committed.
claude
added this to the CI/CD 0: Stabilize the runner milestone 2026-11-09 17:05:57 +00:00
Resources. Runner container: request 50m CPU / 64Mi, limit 500m / 256Mi. DinD: request 500m CPU / 2Gi, limit 2 CPU / 8Gi, plus S3's ephemeral-storage values. Sized for one job at a time on a 2-CPU, 16 GB worker, which is S3's recommended node. A job that goes over the DinD memory limit is OOM-killed inside the pod and fails with a step error, rather than the node evicting the whole runner.
capacity: 1 (was 2).
Timeouts.timeout: 2h, kept below Gitea's ENDLESS_TASK_TIMEOUT from S6 (#39) so the runner, which can say why, fails a long job first. fetch_timeout: 10s.
Shutdown. Chart 0.1.2 doesn't expose terminationGracePeriodSeconds, so the pod keeps Kubernetes' 30 s default. I set shutdown_timeout: 15s and report_close_timeout: 10s so that on SIGTERM the runner cancels running jobs and reports them to Gitea inside those 30 s. Before this, shutdown_timeout was 0. DinD runs as a native sidecar, so it stops after the runner and stays up while jobs are cancelled.
Images.force_pull: false. The three labels are pinned to runner-images:ubuntu-24.04-v26.11.03 and ubuntu-22.04-v26.11.03. On 2026-11-10 those have the same digests as the floating ubuntu-24.04 and ubuntu-22.04 tags, so nothing changes for jobs except that the image stops moving.
Option names were checked against gitea-runner generate-config from the docker.gitea.com/runner:2.0.1 image that's deployed.
Rendered runner config (Ansible render of the role values, then helm template on chart 0.1.2):
The StatefulSet renders the requests and limits above for runner and dind.
Tested: the values above went through Ansible (stub secrets, Helm module replaced by a stub that echoes its values) and then through helm template. gitea.yaml --syntax-check passes, and --list-tasks --tags actions is unchanged: only the runner tasks. Nothing was applied.
Deploy. This comes along with S3, because S3's node-label check is already in the role:
Effect: the runner pod is recreated on the labeled node with the new limits and config. Gitea itself isn't touched.
Risk: the restart kills any running CI jobs, so check the Actions pages first. If the chosen node can't fit a 500m CPU request, the pod stays Pending. In that case lower gitea_actions_dind_cpu_request or pick another node. With capacity 1, jobs queue instead of running in parallel.
Rollback:git revert 91b81e5 in the superproject, then rerun the same command.
Not yet verified: the second acceptance criterion, that a forced pod restart produces a clear job failure, can only be checked after deploy. That's S8 (#41).
— eom-cicd-roadmap
**Committed, ready to deploy together with S3 (#36).** DevOps/ansible-role-gitea@5fb2d9c, bumped in 91b81e5.
**What changed (all role defaults, with comments)**
- **Resources.** Runner container: request 50m CPU / 64Mi, limit 500m / 256Mi. DinD: request 500m CPU / 2Gi, limit 2 CPU / 8Gi, plus S3's ephemeral-storage values. Sized for one job at a time on a 2-CPU, 16 GB worker, which is S3's recommended node. A job that goes over the DinD memory limit is OOM-killed inside the pod and fails with a step error, rather than the node evicting the whole runner.
- **`capacity: 1`** (was 2).
- **Timeouts.** `timeout: 2h`, kept below Gitea's `ENDLESS_TASK_TIMEOUT` from S6 (#39) so the runner, which can say why, fails a long job first. `fetch_timeout: 10s`.
- **Shutdown.** Chart 0.1.2 doesn't expose `terminationGracePeriodSeconds`, so the pod keeps Kubernetes' 30 s default. I set `shutdown_timeout: 15s` and `report_close_timeout: 10s` so that on SIGTERM the runner cancels running jobs and reports them to Gitea inside those 30 s. Before this, `shutdown_timeout` was 0. DinD runs as a native sidecar, so it stops after the runner and stays up while jobs are cancelled.
- **Images.** `force_pull: false`. The three labels are pinned to `runner-images:ubuntu-24.04-v26.11.03` and `ubuntu-22.04-v26.11.03`. On 2026-11-10 those have the same digests as the floating `ubuntu-24.04` and `ubuntu-22.04` tags, so nothing changes for jobs except that the image stops moving.
- Option names were checked against `gitea-runner generate-config` from the `docker.gitea.com/runner:2.0.1` image that's deployed.
**Rendered runner config** (Ansible render of the role values, then `helm template` on chart 0.1.2):
```yaml
container:
require_docker: true
docker_timeout: 300s
force_pull: false
runner:
capacity: 1
timeout: 2h
shutdown_timeout: 15s
report_close_timeout: 10s
fetch_timeout: 10s
labels:
- "ubuntu-latest:docker://docker.gitea.com/runner-images:ubuntu-24.04-v26.11.03"
- "ubuntu-24.04:docker://docker.gitea.com/runner-images:ubuntu-24.04-v26.11.03"
- "ubuntu-22.04:docker://docker.gitea.com/runner-images:ubuntu-22.04-v26.11.03"
```
The StatefulSet renders the requests and limits above for `runner` and `dind`.
**Tested:** the values above went through Ansible (stub secrets, Helm module replaced by a stub that echoes its values) and then through `helm template`. `gitea.yaml --syntax-check` passes, and `--list-tasks --tags actions` is unchanged: only the runner tasks. Nothing was applied.
**Deploy.** This comes along with S3, because S3's node-label check is already in the role:
```
kubectl label node <chosen-node> eom.dev/ci-runner=true
ansible-playbook -i inventories/eom.dev.yml gitea.yaml --tags actions
```
Add your usual vault option.
- **Effect:** the runner pod is recreated on the labeled node with the new limits and config. Gitea itself isn't touched.
- **Risk:** the restart **kills any running CI jobs**, so check the Actions pages first. If the chosen node can't fit a 500m CPU request, the pod stays Pending. In that case lower `gitea_actions_dind_cpu_request` or pick another node. With capacity 1, jobs queue instead of running in parallel.
- **Rollback:** `git revert 91b81e5` in the superproject, then rerun the same command.
**Not yet verified:** the second acceptance criterion, that a forced pod restart produces a clear job failure, can only be checked after deploy. That's S8 (#41).
— eom-cicd-roadmap
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
The runner has no CPU or memory requests or limits and runs two jobs at once, on nodes that mostly have 1 CPU. It uses default job and shutdown timeouts.
Work
capacityto what the chosen node can sustain (likely 1 until a build host exists).runner.timeout,shutdown_timeout,fetch_timeout, and a matchingterminationGracePeriodSecondsso jobs can finish or report during a restart.container.force_pull: falseand pin the three runner image labels instead oflatestwhere practical.Acceptance
Depends on: S2. Label
ready-to-deploywhen committed.Committed, ready to deploy together with S3 (#36). DevOps/ansible-role-gitea@5fb2d9c, bumped in
91b81e5.What changed (all role defaults, with comments)
capacity: 1(was 2).timeout: 2h, kept below Gitea'sENDLESS_TASK_TIMEOUTfrom S6 (#39) so the runner, which can say why, fails a long job first.fetch_timeout: 10s.terminationGracePeriodSeconds, so the pod keeps Kubernetes' 30 s default. I setshutdown_timeout: 15sandreport_close_timeout: 10sso that on SIGTERM the runner cancels running jobs and reports them to Gitea inside those 30 s. Before this,shutdown_timeoutwas 0. DinD runs as a native sidecar, so it stops after the runner and stays up while jobs are cancelled.force_pull: false. The three labels are pinned torunner-images:ubuntu-24.04-v26.11.03andubuntu-22.04-v26.11.03. On 2026-11-10 those have the same digests as the floatingubuntu-24.04andubuntu-22.04tags, so nothing changes for jobs except that the image stops moving.gitea-runner generate-configfrom thedocker.gitea.com/runner:2.0.1image that's deployed.Rendered runner config (Ansible render of the role values, then
helm templateon chart 0.1.2):The StatefulSet renders the requests and limits above for
runneranddind.Tested: the values above went through Ansible (stub secrets, Helm module replaced by a stub that echoes its values) and then through
helm template.gitea.yaml --syntax-checkpasses, and--list-tasks --tags actionsis unchanged: only the runner tasks. Nothing was applied.Deploy. This comes along with S3, because S3's node-label check is already in the role:
Add your usual vault option.
gitea_actions_dind_cpu_requestor pick another node. With capacity 1, jobs queue instead of running in parallel.git revert 91b81e5in the superproject, then rerun the same command.Not yet verified: the second acceptance criterion, that a forced pod restart produces a clear job failure, can only be checked after deploy. That's S8 (#41).
— eom-cicd-roadmap