Skip to content

ci: check images resolve, the stack boots and databases keep their data - #8

Merged
dominics merged 1 commit into
mainfrom
ci/pr-checks
Nov 14, 2026
Merged

dominics merged 1 commit into
mainfrom
ci/pr-checks

Conversation

@dominics

@dominics dominics commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner
🤖 LLM generated description follows

What this adds

Renovate's pull requests here have had no checks at all. This adds one workflow, .github/workflows/ci.yml, with six jobs that run on pull_request and on pushes to main. Most changes to this repository are an image tag or digest moving, so the jobs are about images: does the reference exist, does the container boot, does Traefik still route, does a database still open its data.

It runs on hosted ubuntu-latest runners, not the ones in ci-runner/. The repository is public, so hosted minutes are free, and a self-hosted runner would run a fork's pull request on dagger. No secrets are used: .github/ci/dummy-env writes the .env files from .env.example with placeholders.

Jobs

job what it protects time
compose docker compose config accepts each compose file in the repository (root, immich/, ci-runner/) with a .env built from .env.example. A required variable added to a compose file but not to .env.example fails here. ~5 s
images Each image the root and Immich compose files name can be pulled: the pinned digest exists and has a linux/amd64 build, and the tag beside it exists. A tag that has moved on from its pin is a warning, not a failure, since that is the normal state between Renovate's digest PRs. ~30 s
build Each Dockerfile in the repository builds (today that is ci-runner/Dockerfile). ~1 min
smoke The main stack boots on empty volumes: traefik, redlib, td-replays, garage, ryot-db, ryot, sonarr, radarr, lidarr, prowlarr, sabnzbd, nzbdav, both aiostreams, plex and jellyfin. Each answers HTTP on its published port (the *arrs, Jellyfin and Plex on their health endpoints), no container has restarted, and Traefik's API reports each router, middleware and service as enabled, including the geoblock plugin. Then real requests: :80 redirects to https, www.statio.nz passes geoblock and redirects to the bare name, the AIOStreams route answers 403 from outside the allowlist, and /replays/ reaches the td-replays nginx through the docker provider. ~2 min
smoke-immich Immich's server boots against its Postgres and Valkey and answers /api/server/ping, which it only does after reaching both and running its migrations. ~1 min
upgrade Each Postgres (ryot-db, Immich's database) opens a data directory written by the image on the base commit: start the base image, write a row, swap to this commit's image on the same volume, read the row back. This is the check an empty-volume boot cannot make. ~40 s

Job names are stable and can be made required checks as they are.

Evidence that it fails when it should

Throwaway commits were pushed to this branch and removed again; the runs remain.

change result
Sonarr's digest with one hex digit altered images fails: "the pinned digest does not resolve". smoke fails too, at the pull. (run)
dynamic.yml: the aiostreams router names tailscale-onyl smoke fails twice over: Traefik's API reports aiostreams@file with middleware "tailscale-onyl@file" does not exist, and the route answers 404 where 403 is expected. The other five jobs pass. (run)
#7's change, postgres 17 -> 18 upgrade and smoke fail, see below. (run)

The first green run also only passed smoke-immich by winning a race (the server starting before Postgres was ready and being restarted); a later run lost it. The script now starts the database first, and that job has passed on all runs since.

What it says about the open Renovate PRs

Each PR's commit was cherry-picked onto this branch and run through the workflow. The PRs themselves will show checks once this merges and they rebase.

PR change verdict
#3 valkey:9 digest (Immich) All six jobs pass. Immich's server boots against the new Valkey. (run)
#4 postgres:17-alpine digest (ryot-db) All six pass. The new digest opens the data directory the current one wrote, and Ryot migrates against it. (same run)
#6 traefik v3.7.12 -> v3.7.18 All six pass. Routers, middlewares and the geoblock plugin all come up enabled under 3.7.18 and the four routing probes pass. (same run)
#5 traefik:v3.7.12 digest All six pass (run). It edits the same line as #6, so only one of the two is needed.
#7 postgres 17 -> 18 (ryot-db) Fails upgrade and smoke. upgrade: "postgres:18-alpine does not start on a data directory written by postgres:17-alpine" (job). smoke fails as well (job): the 18 image refuses to start with a volume mounted at /var/lib/postgresql/data and wants it at /var/lib/postgresql, so ryot-db never becomes healthy and Ryot never starts. So #7 needs the mount moved in the compose file as well as a pg_upgrade or dump and restore of the live data; the image bump alone would take Ryot down.

The images job also notes, as warnings, that nginx:alpine has moved on from its pinned digest; Renovate has no PR open for that one yet.

What it cannot catch

  • td-server is not checked at all. Its image is a private package of dominics/td and this repository's token cannot read it, so images skips it and smoke does not start it. The traged@docker router is therefore never built in CI. Granting this repository read access to the package (Package settings -> Manage Actions access) would close that, with no new secret.
  • No real certificates. The smoke run points both ACME resolvers at an address nothing listens on, so CI never asks Let's Encrypt for anything. HTTP-01, Route53 DNS-01 and the AWS_* credentials are untested, and routes are exercised on Traefik's default certificate.
  • Empty state only, apart from the two Postgres volumes. A new Sonarr, Plex or Ryot release that boots clean but mishandles an existing config directory or fails a migration from real data passes here. The upgrade job proves Postgres opens the old data directory, not that Ryot's or Immich's migrations succeed against years of rows.
  • No GPU. Plex and Jellyfin run with runtime: runc in CI, so the nvidia runtime and hardware transcoding are untested.
  • No real upstreams. statio.nz's backend (192.168.1.200:8181), reddit behind redlib, usenet behind nzbdav, TMDB behind Ryot: none is reached. redlib is held only to "nginx answered".
  • sabnzbd is only checked for staying up. A fresh config listens on 8080 while the compose file publishes 8081, so there is nothing to probe from outside.
  • Immich's machine-learning container is not started (it is several gigabytes); its image reference is still resolved by images.
  • A tag paired with the wrong digest passes with a warning. The digest is what Docker enforces, so the deploy gets what the digest says; the tag is only a label.
  • deploy-dagger, tailscale/apply-serve and the traged/ scripts are not run or linted, and neither is the Home Assistant config.
  • The geoblock plugin is downloaded from plugins.traefik.io on each run, as it is on a fresh host. An outage there fails smoke for reasons that have nothing to do with the change.

🤖 Generated with Claude Code

@dominics
dominics force-pushed the ci/pr-checks branch 7 times, most recently from 09a24dc to ed59eda Compare November 14, 2026 14:49
Renovate's pull requests arrived with no checks. Six jobs on hosted
runners now say whether an image bump is safe to deploy: the compose
files parse, each tag and digest exists with an amd64 build, the
Dockerfiles build, the main stack and Immich boot on empty volumes and
Traefik routes as configured, and each Postgres opens a data directory
written by the image it replaces.
@dominics
dominics marked this pull request as ready for review November 14, 2026 14:54
@dominics
dominics merged commit 2c6ebc7 into main Nov 14, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant