Shit! You got deadlines, and managers or customers piling in? Yeah, they don’t pass, but who cares! The code works….probably! Ship it!

This is dumb as fuck.

We have a few non-required checks here and there - mostly as you need an admin to list a check as required and that can be annoying to do. And we still get code merged in occasionally that fails those checks. Hell, I have merged in code that fails the checks. Sometimes checks take a while to run, and there is this nice merge when ready button in GH. But it will gladly merge your code in once all the required checks have passed ignoring any non-required checks.

And it is such a useful button to have, especially in a large codebase with lots of developers - just merge in the code when it is ready and avoid forgetting about things for a few hours and possibly having to rebase and run all the checks again because of some minor merge conflict...

But GH required checks are just broken for large code bases as well. We don't always want to run every check on every code change. We don't need to run all unit tests when only a documentation has changed. But required checks are all or nothing. They need to return something or else you cannot merge at all (though this might apply to external checks more then gh actions maybe). I really wish there was a require all checks to pass and a at least one check must run. Or if external checks could tell GH when they are required or not. Either way there is a lot of room for improvement on the GH PR checks.

Would you look at that - the pipeline is green now! Quick everybody, merge your stuff while it's stable (/s) (sadly a true story tho)

Why write tests when you should be writing more features the business needs now but will never use?

There are definitely ways to run partial testing suites on modified code only. I feel like much of what you're complaining about is an already solved problem.

If you only write tests for things that won't fail, you're doing it wrong. Are you anticipating some other feature coming soon? Write a failing test for it. Did you find untested code that might run soon with a little work? Write a test for it. Did a nonessential feature break while adding an essential feature, let the test fail and fix it later.

a bug? No problem we will just fix it in the next release. loop for eternity.

This just sounds like "my frontend only changes shouldn't be impacted by some dumbass breaking backend two commits ago", which seems reasonable.

Bro just crash the CI because the linter found an extra space bro trust me bro this is important. Also Unit tests are optional.

Eww, no. You're doing tests wrong. The point of tests is to understand whether changes to the code (or dependencies) break any functionality. If you have failing tests, it makes this task very difficult and time consuming for people who need it most, i.e. people new to the project. "Is this test failing because of something I've done? <half an hour of debugging later> Oh, it was broken before my changes too!". If you insist on adding broken tests, mark them as "expected to fail" at least, so that they don't affect the overall test suite result (and when someone fixes the functionality they have to un-mark them as expected to fail), and the checkmark is green. You should never merge PRs/MRs which fail any tests - it is an extremely bad habit and forms a bad culture in your project.

The real problem is merging before waiting for that one slow CI pipeline to complete

It can be finicky to set up and mistakes can be made easily. Often you have to manually replicate the entire internal dependency tree of your project in the checks so that there are no false positive test results. There are some per-language solutions, and there's Nix which is almost built for this sort of thing, but both come with drawbacks as well.

Yeah there are ways to run partial tests on modified code only. But they interact poorly with GH required checks. https://github.com/orgs/community/discussions/44490 goes into a lot more detail on similar problems people are having with GH actions - though our problem is with external CICD tools that report back to GH. Though it does look like they have updated the docs that are linked to in that discussion so maybe something has recently changed with GH actions - but I bet it still exists for external tooling.

One problem is GHs auto-merge when ready button. It will merge when there are still tests running unless they are required. It would be much better if the auto merges took into account all checks and not just required ones.

You're both right. You're both wrong.

• You write tests for functionality before you write the functionality.
• You code the functionality so the tests pass.
• Then, and only then, the test becomes a regression test and is enabled in your CI automation.
• If the test ever breaks again the merge is blocked.

If you only write tests after you've written the code then the test will test that the code does what the code does. Your brain is already polluted and you're not capable of writing a good test.

Having tests that fail is fine, as long as they're not part of your regression tests.

There are two different things mentioned here, which I feel I need to clarify:

First, what you said about merging / creating a PR with broken tests. Absolutely you shouldn't do that, because you should only merge once the feature is finished. If a test doesn't work, then either it's testing for the wrong aspect and should be rewritten, or the functionality doesn't work 100% yet, so the feature isn't ready to get merged. Even if you're waiting for some other feature to get ready, because you need to integrate it or something, you're still waiting, so the feature isn't ready.

At the same time, the OP's point about tests being supposed to fail at first isn't too far off the mark either, because that's precisely how TDD works. If you're applying that philosophy (which I personally condone), then that's exactly what you do: Write the test first, checking for expected behaviour (which is taken from the specification), which will obviously fail, and only then write the code implementing that behaviour.

But, even then, that failing test should be contained to e.g. the feature branch you're working on, never going in a PR while it's still failing.

Once that feature has been merged, then yes, the test should never fail again, because that indicates a new change having sabotaged some area of that feature. Even if the new feature is considered "essential" or "high priority" while the old feature is not, ignoring the failure is one of the easiest ways to build up technical debt, so you should damn well fix that now.

I concede that on a feature branch, before a PR is made, it's ok to have some failing tests, as long as the only tests failing are related to that feature. You should squash those commits after the feature is complete so that no commit has a failing test once it's on master.

(I'm also a fan of TDD, although for me it means Type-Driven Development, but I digress...)

• You write tests for functionality before you write the functionality.
• You code the functionality so the tests pass.
• Then, and only then, the test becomes a regression test and is enabled in your CI automation.
• If the test ever breaks again the merge is blocked.

I disagree. Merging should be blocked on any failing test. No commit should be merged to master with a failing test. If you want to write tests first, then do that on a feature branch, but squash the commits properly before merging. Or add them as disabled first and enable after the feature is implemented. The enabled tests must always pass on every commit on master.

gitlab has a feature where you can set it to auto-merge when and if the CI completes successfully