Six apps tested live, a stuck-Failed bug, and a docs and UX sweep
We ran six self-hosted apps on production as throwaway apps (all work, with recipes), fixed a bug that left healthy apps marked Failed, and swept the docs and UI.
This morning we pointed PicoPaaS at six popular self-hosted apps, ran each one as a throwaway app on production, and wrote down what it took. All six now work, and each has a recipe in the docs. We also found and fixed a real platform bug along the way.
Six apps, tested live
- Vikunja (to-do and kanban): the public image runs as is. The Postgres connection is CA-verified, and a negative test with the wrong CA failed with "unknown authority", so the verification is real. Attachments go to the bucket. About 27 MiB of RAM.
- Umami (analytics): also as is, but memory is tight: roughly 166-254 MiB idle and a peak near 328 MiB during first-boot migrations, against a 384 MB limit. It stores one row per event, so use it on low-traffic sites, and change the default
admin/umamilogin right away. - Planka (kanban): needed three fixes for our fixed UID. Its database connection is encrypted but not CA-verified, and we say so in the recipe. Its fair-use license bars running it as a hosted service for third parties, so you can run it yourself on PicoPaaS, but we will never offer it as a managed service.
- ntfy (push notifications, on Postgres): needs a small wrapper image. Our first published wrapper forgot
CMD ["serve"]and just printed help text and exited. PicoPaaS has no command override, so bake the start command into the image and test wrappers with a baredocker run. The retest passed, including the negative TLS test, at about 15 MiB. - IT-Tools and Excalidraw: static wrappers on nginx-unprivileged, about 20 MiB each.
These are recipes, not one-click templates yet; whether to promote them is a decision still pending.
A bug the tests found
An app whose crashed revision had been replaced by a healthy one could stay "Failed" for up to an hour. Two causes: status reads picked up errors from old or terminating pods, and errored apps were only rechecked after a backoff. Now status looks only at the current revision's pods, and errored apps get a quick health probe each sweep.
Our security review caught a subtle trap in the first version: a rejected deploy, where the old version keeps serving, must not be "recovered", or we would mark an image that never ran as known-good and defeat auto-rollback. That is fixed and tested. The fix is rolling out with today's deploy.
Docs and UX sweeps
Every API route is now documented, about 38 missing env vars were added, 8 contradictions were fixed, and the README has a docs map. On the UI side: a "Resume to retry" message that had no Resume button now has one, a Failed banner that never refreshed now polls, New app and Add variable are labeled buttons, and long hosts wrap at 390px. A couple of layout proposals are waiting on Chas.