On 12.07.20 14:33, Marius Bakke wrote: > Jonathan Brielmaier <jonathan.brielma...@web.de> writes: > >> On 12.07.20 03:44, Christopher Lemmer Webber wrote: >>> Commit 6680880f9b8dceb4f2f3f91bd2b13c659b53835e pushed out a new version >>> of Guix, and it looks like it wasn't possible to build new systems from >>> that because the filename for the "Berlin ACL key" changed. (Or at >>> least, I couldn't run "guix system vm".) >>> >>> I pushed out a "fix" for this. I hope it's ok. >> >> Thanks for the fix. >> >> As I ran into all those little errors with `guix pull` this week-end, I >> wonder if we can do better. > > This particular change broke 'guix system', not 'guix pull'. Which is > equally bad of course, but a different kind of beast entirely. > > Are you referring to something else?
It started with d283bb960f927dd5f7bb8b96bc697221e4e8ad39 and it could be `guix system` which failed. Then it's a different case, because testing `guix system` is working, is more difficult. > >> So maybe some pre-checkin CI which tests that a commit/commit series >> doesn't break `guix pull`. What do you think? Is this doable? > >> I find those little errors pretty annoying as they seem to be avoidable >> through technical counter measures... > > One possible solution that has been discussed before is to have the CI > continously merge master to a 'stable' branch when lights are green. > There are quite a few challenges to solve with that approach though. > > We could make the pre-push hook run 'guix pull' and 'guix system build' > but it will quickly get annoying. A server-side hook for the same would > be less annoying, but would have a hard time if someone accidentally > pushes a full rebuild. That's a point. So my idea was to do it at least for all changes of Guix itself. Like updating the guix package, new po files, Makefile changes etc. I think this could lead to quite an improvement. > In practice there will always be problems that cannot be caught in an > automated way. I hope such breakages are rare, but from your message it > sounds like there were many problems just this week-end? I counted two or three this week-end connected with the renaming of the public key of the berlin build farm, so a "trivial" change...
