I am yet to see a real value to use JJ. Those kind of built-up posts I am seeing about JJ on HN tend to push me in the opposite direction.
1) do your test in your current branch, it's related to what you are doing. 2) commit, co master, co -b newtestbranch, PR it, merge in your branch after. You're going to squash rebase at the end anyway...
I have experienced this exact scenario hundreds of times using git over the past two decades. I’m in the middle of some work, I realize I need to task switch to do some other work, then the work that was mid-flight should depend upon that interim thread of changes.
In git it is doable but it is a process, and one that is easily derailed (oh god a chain of failed automatic rebases) that requires keeping state in my head. In jj it is so trivial it doesn’t even register that I did something interesting or special, except when I remember how awful it would have been doing that in git.
> You're going to squash rebase at the end anyway
No? I’m not?
git lets you create as complicated a commit tree as you want, and lets you do merges so you can have a separate commit where multiple changes come together, but realistically it really only works if you use it for a rebase workflow in your branches, squashing all changes into a single commit each time. Merge is an outlier when you need a more powerful tool to join branches but want to keep both of the originals still intact. But that shouldn't be your everyday workflow or you're guaranteed to end up with an unreadable, untraceable, and unmanageable commit graph.
What’s actually the case is that your workflow seems best with squash merging. There are a lot of project and teams where this is not the case.
All of these workarounds just… don't exist in jj, nor do they need to.
> 1) do your test in your current branch, it's related to what you are doing.
The article describes a scenario where you “feel reasonably assured that green CI will mean you’re on the right track, and you’ve been removing parts of the old parser as you introduce the new”. But this one parser feature you discovered was not under test. “The original parser is half-taken to bits,” so perhaps you already deleted the implementation of that feature, not knowing it was depended on. If the implementation were missing, of course you couldn’t test it on your current branch.
Okay, let’s say you didn’t delete the implementation yet – but the implementation calls some of the methods you already modified. If you write a test on your current branch, it would only show that your already-modified parser passes the test. You couldn’t be confident that the expectations in the test you wrote actually match the behavior of the old parser. If the old parser acted differently from your tests (even if it was due to a bug), you need to know about that so you can update callers of the old parser to not rely on that behavior any more.
To be confident that your new test accurately describes behavior you need to support, you need to run the new test against the “develop” branch.
> 2) … You're going to squash rebase at the end anyway...
Why would you assume that? Some teams prefer to express changes as a series of independent commits when possible, and they merge PRs while keeping those individual commits. My current team is one example. On such teams, keeping a Git feature branch’s history organized is a real need.
If you wonder why a team wouldn’t just squash rebase, I think the goals of keeping commits separate on the main branch are to make debugging tools such as `git log`, `git blame`, and `git bisect` more useful.
*actually work.
No point using `git bisect` if you switch to a commit that literally doesn't even compile.