[SamBam]

Hmmm, I've always believed that no commit should break a build, even if you're committing the fix right after. Otherwise you're going to cause problems for `git bisect` or other practices of going through the history to find where a problem may have started.

Do other people commit breaking tests and then fixes?

[Bartweiss]

I think this depends quite a bit on what other contributors are doing - it's one of those cases where several approaches are acceptable, but inconsistency is not.

"Commits should always build" is one doctrine I've seen. As you say, it makes bisecting and other error-analysis approaches easy. On the other hand, it risks either having large, opaque commits, or adding overhead to make intermediate commits build - possibly with flawed/meaningless behavior when they do.

Another is "the trunk should always build". In that case, you'd just squash branch commits down to logical groupings that are easy to analyze, whether or not things build. You can bisect on the trunk, but lose all guarantees about state on branches.

Finally, I've seen variations on "no commits that break the product", "no commits that make things worse", or "no committing failing tests without subsequent fixes". In this case, you can't generally commit broken builds, but can specifically add failing tests. The first rule just means "adding failing tests is ok", the second means "converting runtime bugs to failing tests is ok", and the third means "write your test and fix, but split (and ideally tag) the commits". All of these break bisect, but they guarantee the project itself won't become more broken from commit to commit, and they can help with other forms of reasoning about where bugs first occurred.

Every approach there seems viable if you stick to it. If there's no established practice, I suppose the best choice would be based on what sort of work and debugging is most likely to apply.

[monsieurbanana]

I might be missing something basic here. Isn't the "no commit should break a build" impossible to enforce on a codebase where you need to push a commit to run the tests?

Something where you can't test locally, like when testing on multiple architectures or when the tests just take too long for a laptop.

[SamBam]

In my workflow, that would be in a feature branch, and exploratory branches can certainly break, but before I made a PR I would rebase my changes such that none of the commits broke the build.

It's also a different situation. The original one is "I've made a test that shows a problem." Your example is a surprise "I don't know whether this will pass my cloud-based tests." I would edit my branch if I had a surprise failure, since my initial code clearly wasn't correct.

[claytonjy]

I do this, but not in a way then end up on master. It's a driving force behind my preference for squash-and-rebase merge patterns.

A good bugfix PR is often two commits then: one with a test to catch the breakage, another to fix it so the tests pass. Reviewers can see the failing-then-passing CI job logs, so if they agree your test catches the bug, they have additional CI-automated validation your fix worked.

Then as long as you squash when completing the merge, you get the best of both worlds.