Skip to content

ci(renovate): run on the unprivileged in-cluster runner; never automerge EMQX - #4585

Closed
dapperdivers wants to merge 4 commits into
mainfrom
ci/renovate-unprivileged-runner
Closed

dapperdivers wants to merge 4 commits into
mainfrom
ci/renovate-unprivileged-runner

Conversation

@dapperdivers

Copy link
Copy Markdown
Owner

Summary

Renovate now runs in-cluster on the unprivileged ARC runner set from #4584, and EMQX is never auto-merged.

Renovate on dapper-cluster-unprivileged

  • runs-on: dapper-cluster-unprivileged with container: ghcr.io/renovatebot/renovate:44.216.1@sha256:…. In ARC kubernetes mode, the steps run in their own job pod.
  • renovatebot/github-action is dropped because it needs Docker, which kube-mode runners don't have. The workflow now runs the renovate binary directly. It no longer does a checkout, since Renovate clones the repo itself.
  • The image is pinned by digest (Renovate bumps it) instead of renovate-version: latest. The version dispatch input is removed.
  • RENOVATE_BASE_DIR=/tmp/renovate.

Verified with a dry run from this branch (dryRun=full, run 41946456740):

  • Every step succeeded, ending in Repository finished, with no errors.
  • The job pod …-workflow ran under SA default with automountServiceAccountToken: false, as uid 1001 / gid 0 with allowPrivilegeEscalation: false. Its only volume was the job work dir: no Talos secret and no token.
  • The runner logged BrokerServer/ResultServer websocket timeouts to GitHub (live log streaming). They didn't affect the job, and its lease kept renewing.

EMQX: automerge disabled

That same dry run planned Would merge PR #4072 (emqx 6.2.2 ➔ 6.2.3). That's the patch bump that wedged the broker for 8 days in August (rolled back in #4070), because the operator can't roll even patch bumps. It had sat unmerged only because Renovate barely ran (about every 7 hours). Today's schedule fixes made it about to go through. #4072 is now closed. This PR removes /emqx/ from the standard-apps automerge list and adds an EMQX deny rule (broker + operator).

Validation

  • renovate-config-validator --strict: valid
  • zizmor: clean

…tainer

runs-on dapper-cluster-runner with container: ghcr.io/renovatebot/renovate
(digest-pinned). In ARC kubernetes mode a job container runs as its own
pod, so Renovate never shares the runner pod's Talos os:admin mount; the
job pod gets the namespace default ServiceAccount (no RBAC).

- drop renovatebot/github-action (needs docker, which kube-mode runners
  don't have) and the unused checkout; run the renovate binary directly
- pin the image by digest instead of renovate-version: latest
- RENOVATE_BASE_DIR=/tmp/renovate (job workdir is owned by the runner uid)
/emqx/ was in the standard-apps automerge list, so once Renovate ran
reliably it was about to merge #4072 (emqx 6.2.3) - the same patch bump
that wedged the broker for 8 days in August (#4070). The operator can't
roll even patch bumps; EMQX carries all Zigbee/HA control traffic. Move
EMQX (broker + operator) to the deny block.
@dunce-bot

dunce-bot Bot commented Nov 5, 2026

Copy link
Copy Markdown
Contributor

Note

No rendered changes.

konflate 0.6.6 · rendered 1a8c5b4 · full diff →

@dapperdivers

Copy link
Copy Markdown
Owner Author

Closing: Renovate stays on GitHub-hosted runners. Measured on the unprivileged in-cluster runner (dry runs after the 11-05 egress outage cleared): cold 370s / warm 342s total, versus 140–160s on ubuntu-latest. Lookups are WAN-latency-bound and the job pod needs 1–2 min to pull the image and mount the cache. The cron trigger is GitHub's either way, so there's no reliability gain, and an ephemeral GitHub VM isolates better than a pod on the cluster network. The EMQX automerge block moved to #4591.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant