Every CI job names its runner image, Ubuntu 26.04 where a trial ran it, and stops at about three times its usual length¶
Decision 796 · accepted · 2026-10-06
The trial¶
A temporary workflow ran the critical jobs on ubuntu-26.04 once, with no cache, so every
dependency was fetched and compiled on the new image (run 37483437510, image
ubuntu-26.04 version 20260927, Ubuntu 26.04.1 LTS, kernel 7.0.0). Every job passed:
| job | what it ran | result |
|---|---|---|
| umbrella | erlef/setup-beam (OTP 28.5.0.5 built for ubuntu-26.04), Zig, inotify-tools and bubblewrap with the AppArmor sysctl, scripts/dev-up in Docker, mix compile --force --warnings-as-errors, migrations, troupe_protocol's suite, core's sandbox and file-watch suites |
passed, 3.0 min: the protocol's 185 tests and core's 12, the sandbox's among them, which flunk without a working bubblewrap |
| TUI | the two locks agree, mix compile --warnings-as-errors, mix troupe.cli.reference --check |
passed, 2.3 min |
| GUI | tokens:check, icons:check, typecheck, build, test |
passed, 1.6 min |
| native TUI | native.yml's linux_x86_64 musl build (Burrito and its ERTS download) and its headless Fake smoke run |
passed, 5.7 min |
| pages and chart | setup-python 3.12, mkdocs build --strict, helm lint, kubeconform in Docker |
passed, 1.0 min |
| VS Code | typecheck, unit tests, the suite inside a downloaded VS Code under xvfb-run, pnpm package |
passed, 0.9 min |
What runs where¶
- Every job that said
ubuntu-latestsaysubuntu-26.04. Named rather than left to the label, so the move happens on this commit and not in the middle of some release during GitHub's rollout; a run never mixes the two images (the umbrella jobs share one Mix cache keyed onrunner.os, which isLinuxon both); and the next move is again somebody's choice rather than GitHub's. - The cluster suite runs on
ubuntu-24.04. kind and Cilium lean on the kernel more than anything else here, and the trial did not run them: 25 minutes of a runner and the five images. Moving it is changing its label on a branch and running the nightly there (gh workflow run nightly.yml --ref <branch>) before that change merges. - Labels already named stay. The TUI's jobs (
ubuntu-24.04; its compile passed on 26.04 too, so it can move when someone wants), the daemon's and the TUI's native builds (ubuntu-24.04and-arm: the daemon's tarball carries the build host's ERTS, so the build host's glibc is the floor for whoever installs it), the desktop app's Linux build (ubuntu-22.04),live.ymlandquick-start.yml. windows-latestandmacos-latest(the VS Code extension's matrix, the desktop app's Windows build) are not part of this: GitHub's notice is about Ubuntu.
Timeouts¶
Every job that runs on a runner has timeout-minutes, at about three times what it
usually takes and above the slowest successful run seen, from the last twenty successful
runs of release.yml, nightly.yml and prerelease.yml (and fifteen of the light
workflows). A runner that hangs fails its job in minutes instead of GitHub's six hours.
A matrix that spans platforms gets one per target: timeout in native.yml's target
lists, read as timeout-minutes: ${{ matrix.timeout }}.
A timeout that fires on a slow but healthy run is a re-run, not a reason to raise it: the
Windows daemon build that spent 29 minutes in setup-zig passed, and at 15 it would have
been run again. A value goes up when a target's usual time does.
The ACP agent test¶
The fake agents were a bare elixir with a -pa for each of the build's 58 ebin
directories, each looked in first for every module the script's compile loads. The agents
now speak JSON through the standard library's JSON and take no -pa, and the module's
timeout is 60 seconds, above the 38 its own waits add up to.
Governs: .github/workflows/