[dcow]

I’m not talking about the author, specifically or criticizing them directly. I’m lamenting how often I encounter people putting intense amounts of effort into arguing about how git should be used. My hypothesis is that if people understood their tools better before getting all evangelical about forcing their entire eng team to work in their style, others wouldn't have to spend so much time trying to convince them that other styles are exist and are valid.

Compare with a construction site: there are obvious uses for tools and there are less obvious uses and I haven't been on a single one where tools are only used specifically in one way for the entirety of the project. They’re used dynamically by different people with different experiences in order to complete the project. Nobody forces other workers to use a specific grip when holding their hammer…

[djur]

This isn't "how to use a hammer", though, it's more like "how to evaluate the completion of work", which is definitely something working teams have to agree on. git itself is almost irrelevant to the discussion, other than its native feature set having some influence on the options.

[morelisp]

If this was just about team review workflow I wouldn't expect to find "easier to rollback" and "don’t need to wait on a particular change to be merged" to be on the list of upsides, as these are properties of the repository structure and not the review workflow.

[Aeolun]

> "easier to rollback" and "don’t need to wait on a particular change to be merged" to be on the list of upsides

Can you explain how these are not just facts? We’re not using stacked branches because it has enough downsides, but these are indubitably upsides of stacked branches.

[morelisp]

Stacked branches tend to produce repository structures that make this easy, but so do lots of other development practices. It is not related to "how to evaluate the completion of work", it's entirely about how you use the hammer.

[johnfn]

I'm not understanding how stacking PRs is inconsistent with knowing git well.

[morelisp]

I can't be certain what dcow is referring to but there are a few things that point to confusion, or maybe an overambitious attempt at simplification.

For example, Stacking PRs keeps the author unblocked. Authors don’t need to wait on a particular change to be merged before starting to build something on top of those changes.

This will fundamentally be up to the author's git skill whether or not they are presenting the PRs to reviewers / mergers as stacked. If they're skilled git users there's little to no cost presenting them one at a time and keeping the not-yet-PRd branches fresh. If they're not skilled git users, they have no hope of managing multiple PRs effectively only be virtue of presenting them all at once.

Or: Since stacking PRs allows you to create a DAG of dependent changes, this natually allows you to manage code changes that need to be submitted in a particular order.

Assuming this is in a single repository, a fast-forward-only commit policy alone ensures this.

Or: stacked PRs use branches, and can have multiple commits in a single atomic change; stacked commits use a single commit as the unit of atomic change.

Stacked PRs usually use branches, but stacked commits also still use at least one branch and could have more. In both the commit is the unit of atomic change, because that's what a commit is.

I'll also add I find the entire language around "branchless" workflows in git confusing, not just from this author. There is no such thing; considering one special branch as "not a branch" or your local and remote and someone else's remote as the "same branch" just because they have the same name is a holdover from older/other VC tools. We don't do any new git users a pedagogical or practical favor by clinging to that view.

[djur]

> This will fundamentally be up to the author's git skill whether or not they are presenting the PRs to reviewers / mergers as stacked. If they're skilled git users there's little to no cost presenting them one at a time and keeping the not-yet-PRd branches fresh.

On multiple teams I've worked on, this has been explicitly discouraged because reviewers often want to see the changeset in as much context as is available, including future changesets.

[dcow]

What's better context than presenting all the related atomic commits at one time in a single pull request so you know exactly which related set of commits are being added and for what reason? Slap a merge commit on and you can revert the whole thing! But your commits are a mess, and I want to see clean ones. That is what interactive rebase is for and what attentive project contributors would do in order to clean up before or during the merge process. Why is it fair to ask someone to change syntax formatting to but not cleanup their commits? Because the latter is a "big ask" for someone who doesn't know their tool.

[djur]

Then we get back to the problem of not having independent review, CI status, and deployability for each individual changeset. There's no amount of git competency that changes the fact that GitHub PRs operate at the branch level, not the commit level. I've actually done exactly what you say many times -- interactive rebase to clean up the commit history -- but the end result of that process is the creation of multiple "stacked" branches, solely so I can submit multiple PRs and conform with the team's expectations.

[dcow]

> There's no amount of git competency that changes the fact that GitHub PRs operate at the branch level, not the commit level.

Most all CI tools support per-commit checks with simple configuration/scripting of their execution DSL/API. In GitHub, push actions already works this way. You can do the same thing on any branch after a PR is opened: on synchronize you can iterate over all the updated refs, run checks, and publish results to the commit status API https://docs.github.com/en/rest/commits/statuses, and https://docs.github.com/en/pull-requests/collaborating-with-....

[arxanas]

git-branchless author here. By "branchless", I mean literally using detached HEAD as the primary state of development. If you're using tools like Gerrit or Phabricator, you never have to explicitly make a branch to get your code merged. If using GitHub, then branches are unavoidable, but it can be nice to do branchless development as part of rapid prototyping (see https://github.com/arxanas/git-branchless/wiki/Workflow:-div...).

Another term you could use is "anonymous branching". This is not technically accurate in the above workflows, but it captures the essence pretty well in Git, more so than "branchless".

[morelisp]

But when you call it "anonymous branching" you lay bare that the only advantage is that you don't need to name your work "branch", and meanwhile you have a workflow that's needlessly incompatible with most other git tooling.

In particular, since this is the one I see usually called out as a benefit of branchless:

Stock Git does not have good ways of rebasing a sequence of branches.

A sequence of branches can be rebased by rebasing (or otherwise rewriting) the longest one (the only one you'll need locally) then pushing the individual commits in the current branch to the remote under any relevant branch names. This doesn't take zero time, but with good git UIs it will take less time than remembering `git move`, and it's not especially hard to do with the stock CLI either.

[arxanas]

> But when you call it "anonymous branching" you lay bare that the only advantage is that you don't need to name your work branch

Sure, I'm only responding to what you were saying about "There is no such thing; considering one special branch as 'not a branch'". There is such a thing in that there is no branch involved in the detached HEAD state. It's not some kind of Git misunderstanding. I think you might be referring to trunk-based development and always building off of the remote main branch instead of having your own local copy, which is unrelated to being "branchless", for the reasons you stated.

For many people (particularly those on Github!), a branchless workflow won't help, so you're free to not use it. In my opinion, it's a workflow that is better compatible than stock Git with code review tools like Gerrit and Phabricator.

I personally argue that anonymous branching is useful even in some branch-based workflows. Mainly, if Git branches are so lightweight to use, why do we also use the command `git stash`, instead of just always creating a new branch for our temporary work? One benefit of anonymous branching is that it consolidates these workflows in a convenient way. Some people don't stash changes or feel that branching in those cases is heavyweight, so anonymous branching doesn't help them at all.

> then pushing the individual commits in the current branch to the remote under any relevant branch names

If I'm understanding correctly, every time you rebase the longest branch, for each commit in the branch, you would manually run e.g. `git push <commit> origin:my-branch-name`? That seems like it would take a lot of time to me. Is the tacit assumption here that you don't have a lot of commits in your branch, so this doesn't take a lot of time?

[morelisp]

> There is such a thing in that there is no branch involved in the detached HEAD state.

But at the very least there's still an existing remote branch, which is the ultimate merge target, for example - perhaps even multiple. Since we're talking about coordination, there's also the actual state at the remote vs. my view of the remote vs. my teammates' views of the remote. But we don't think about this too much, because we can synchronize easily as long as the remote branch has a name. Taking the name off the branch makes it more difficult to do anything with someone else's work other than merge it, e.g. pass a changeset back and forth or hand over a half-complete task.

> why do we also use the command `git stash`, instead of just always creating a new branch for our temporary work?

I have no idea actually, because I don't. I haven't run git stash manually since I learned about autostash, and even before that it wasn't for temporary work but for changes I decided I wanted somewhere else only after writing them. Temporary work mostly does get a branch (usually as a new commit on top of my current local branch).

> If I'm understanding correctly, every time you rebase the longest branch, for each commit in the branch, you would manually run e.g. `git push <commit> origin:my-branch-name`?

Close, except I'd scroll down the list of commits and type `P o RET TAB branch-name` (or something to that effect, it'd be slightly different if I have several remotes). This would take perhaps ~0.5 seconds per branch and not require me to context switch to a terminal.

It does take longer for people using less powerful clients, but virtually all do provide some way to do it even it means a few right-clicks on a menu and clicking some "OK" buttons, and there's value in everyone using analogous operations. And the people using those clients usually can't reliably recover from a large class of "git broke" mistakes, assistance from git-branchless or not, so handing them a CLI and a prayer is out of the question. And even the "long version" is pretty negligible (like, maybe 10 seconds per branch?) compared to everything else you should do to make your changeset approachable for review.

[dcow]

That’s not what I’m saying at all. My comment is a meta comment about why we even need to write grand defenses of these practices in the first place. I’ve been “stacking PRs” forever… when it makes sense to do so. There’s no golden git usage pattern. I’ve seen the industry all the rave about gitflow. And then it wasn’t cool anymore with tags being the preference. Now it’s cool to just forgo branch names entirely? I’ve watched teams fight over whether every PR should be exactly one commit because they don’t realize that you don't always have to squash merge if you want two unrelated changes pulled in at the same time—just keep the commits clean and revert doesn't care. My point is that if you know how to use the tool none of this really matters and all seems quite distracting. I no problem with people sharing what works for them, that’s great (as is this article, abstractly).