Git Branches Have No Rules
wizardzines.com
wizardzines.com
Just speaking from the perspective of someone that's used git since nearly the very beginning (and CVS and Subversion before that), Git's strong adherence to the Unix philosophy ("give enough rope to hang yourself") was definitely a breath of fresh air compared to the restrictive systems.
It certainly is. There are two things that a team needs to do to use git: 1) Team members all learn git 2) A few conventions are agreed upon to facilitate collaboration
People use #2 to replace #1. And that's why we've had to sit through a neverending flow of blog posts obsessing over git workflows.
Git is flexible to whatever your business workflow is, play around with it a bit, choose something, then make changes when it's not working.
The topic fundamentally does not need this many explainers. But if you're supplanting understanding the tool, it does.
The standard concept that I remember is 'a branch is a pointer to a commit', which still feels accurate
People say things like "a branch does this" when they mean "We have rules on the git server that enforce this", or when they mean "We have a convention to do this, you have to keep track of it". One word to describe this is synecdoche - a part represents the whole or vice versa. Some part of the system does something, or some part of the PROCESS does something, and so people say "git does this thing".
Imprecise speech is shorter, but it can waste a great deal of time in the long run.
But yeah fixing up work and cleaning up a branch prior to merging is a feature not a bug. Git got it right.
A commit SHA represents the entire DAG up until that point, not a given changeset. So in your example, branch B is still a pointer to a commit, and branch A is a pointer to a different commit. Where is the convenient lie?
The method I had in mind is an interactive rebase. Assume there are n > 5 commits on my branch, choosing 5 arbitrarily. If I run “git rebase -i HEAD~5” then use the “drop” command on the third-most-recent commit, I do end up with a different DAG. The change is subtractive and not just the removal of the current HEAD.
This kind of history change is definitely a bit less common but not obscure by any means. I’ve rebased to drop commits on my own feature branches many times.
I would argue that is still not a convenient lie of "a branch is a pointer to a commit". Rather, the rebase command is changing what the branch is pointing to, and git tries to make that clear.
After that rebase command, there is a message like `Successfully rebased and updated refs/heads/yourbranch`, showing that the branch named yourbranch is now pointing at something new.
Your original comment:
>Consider the case where I create a new branch B from A then remove a commit. The branch is now more like a "a pointer to a commit on a DAG" where branch B points to a commit on a different DAG than branch A.
My counterpoint taking this to the extreme:
>Consider the case where I create a new branch B from A then run git reset --hard SOMESHA1
In my example, again the new branch is pointing at a completely different commit--because I am repointing the branch.
Yes, git currently stores deltas between commits in its pack files, but that's an implementation detail to save space. Importantly, it's not how git was originally written, but was added at a later time (although, still a long time ago now).
Edit: https://git-scm.com/book/en/v2/Getting-Started-What-is-Git%3...
No matter what convenient lie you use in learning it will trip up someone, sooner or later.
> Unix philosophy ("give enough rope to hang yourself") was
That and RTFM are probably reasons why *Nix will always be a bit player when it comes to anything that has contact with people.
And, yes, Android runs Linux, so it doesn't count as an OS. I wouldn't call JVM or Spider Monkey my OS either.
Git gives you the tools to allow any member of the team to reconcile those histories cleanly for others.
Forcing everyone to be in sync all the time also means lots of merge conflicts.
Git allows for everyone to drift in their own way and provides tools for them to reconcile that drift in a way that others can integrate easily.
Until you embrace the distributed nature and start using tools to manage your history (rebase & reset) , you will suffer with git.
I cringe when editors refer to "rebase & reset" as "weird".
Rebase and reset are the primary power tools of git. Along with reflog, revert, interactive-add .
The worst repositories I've seen are people who `git merge` 20x a day to stay "in sync" and create labyrinthian change logs that are impossible to parse.
Of course, if you are working on a highly distributed project, like the Linux kernel, many of the Git tools are absolutely invaluable. But there's no reason to bother with them if you're not.
Keeping the history clean is more important than just syncing files.
With a repository server that supports merge requests and feedback from CI/CD tooling on commits/push you can motivate people to use merges rather cleanly.
If you add to that very simple rules like
- create branches for every feature
- commit often
- push often
- merge only via merge/pull requests
then even people new to Git can handle it just fine.
Merge conflicts happen when multiple people work on the same file and try to merge their different status. Imho, code bases where multiple people work on the same file at the same time have a more fundamental problem than Git.
Honestly, I still struggle to see the value in rebasing. Maybe because I haven't used it extensively. Superficially, to me at least, it does the same as a "git push -f".
I am curious to hear why it is supposed to be superior, especially with modern CI tools like Github and the likes.
> I still struggle to see the value in rebasing. Maybe because I haven't used it extensively. Superficially, to me at least, it does the same as a "git push -f".
Rebasing is completely different from pushing. Rebasing changes your history. You can do whatever you want with your history: rebasing, creating personal branches and moving them around, etc... and no one will know. Only your pushes matter to the rest of the world, or alternatively when others pull from your repository. "git push -f" means you don't want to play along with whatever is in the repository you are pushing into, it is usually bad as it may desynchronize it from everyone else's fork. You absolutely don't need to do a "git push -f" with rebase, in fact you really shouldn't.
And none of your rules prevent rebasing. The commit/push often rule make the need for rebasing less likely, but if a dev did a rebase, no one will notice, it didn't leave his own machine. He mostly did that to make the commit he pushed less awkward.
Have you ever squashed down a branch before merging back into main? Then you've used rebase.
The purpose of source control (not just git) is to keep a useful version history of your project. If you don't need to know where/when a commit originally branched off from, rebase can keep a cleaner history than merge.
Even rebasing a branch off the latest commit just to merge back in can be useful.
In popular Git ecosystems like Github you have the option to add a merge commit. To me, this also helps understanding the history of the repo. Would the rebasing be equally visible?
Git is a powerful data structure manipulation tool. Use it as such and harvest that power. After all, what you're working on is just a graph of repository states - and you can attach any meaning you find useful to those states.
If the history is important, as you assert, then a tool for rewriting and improving the history is also important.
Do you put manual work into rebases? It sounds like a lot of effort and time when collaborating.
Git rebase preserves a linear history . Keep a reflog on the server (github has one built in) for cases where someone has screwed up the history. and secure shared branches to avoid clobbering it.
How do you specifically investigate the history with rebase and 10+ developers? Do you manually look into commits on the branch, or go by a single file history, or look at merge/pull requests?
Personally, I prefer looking at the context of a file change within a merge/pull request. The specific merged/pull request can be read from the merge commits. And in case a specific file history is of interest, I can backtrack from that file's associated commits.
There are tons of merge commits in the Linux repository, but no useless ones. Note how in its nature Linux process also does implicit rebasing of applied patches when they land in subtrees.
How does one learn how to (1) think in this way (which necessitates not only abstract description but multiple concrete examples) and (2) what workflows and commands are necessary to effectively implement this?
1. a ref (branch or tag) that builds (compiles & passes all tests including UAT)
2. every shared ref has a clean revision history that allows for graceful change management ( read changelog, fork, integrate, revert, A/B test)
Any ref shared by a team member (e.g. git push) must meet those two qualities. Team members must reconcile refs on their end (e.g. rebase) before they can submit a new ref meeting those expectations.
This way you are not prescribing a workflow (e.g. git-flow). The team needs to agree on that to their software delivery pipeline requirements.
In order to deliver these two expectations, they need to understand refs (pointers to a digraph) and how to manage them (rebase). tell them to RTFM the git manual on those topics.
Just like memory management, pointers, algos etc these skills are essential . They are not a "nice to have" to build software.
This shouldn't be confusing, but it is because that's not a natural way to think about revision control, so we instead have to relitigate it all the time for people that haven't learned git's obtuse naming conventions.
The concept “a branch is a pointer” isn’t a confusing. However the practice of managing and predicting the behavior that pointer is complex, error prone, and full of nasty edge cases.
I think people want branches to just do what they want. They complain that branches are confusing, but if you ask them to define how a branch should work, they won't have a complete answer.
There’s two problems here. One, branch tag behavior is genuinely confusing. Two, branch tags are intrinsically complex (which is why they’re confusing).
Git is intrinsically complex. Version control is a hard, complex problem! Pretending that it isn’t complex just results in a lot of confused users and a lot of unhelpful people who think if they yell “it’s actually simple” just a few more times everything will be ok.
I think git is pretty close to the most intuitive interface possible, but the problem git solves is hard, as you say. Even the best interface, programmed by God, might not make this easy.
Mercurial has about as intuitive an interface as is possible IMO. There's still all the intrinsic complexity of distributed version control, but the verbs are meaningful and if I don't know how to do something, knowing the verb is sufficient to use the built-in help to learn how to do the thing I want to do.
The concept of branches as the entire DAG and not just a commit also better matches how I've seen version control workflows be used in practice across multiple teams and companies. Add in rewriteable history, like what Git allows or Evolve enables, and you're golden.
It’s very simple to use and understand. It’s not perfect or anything. But Git is pretty damn far from intuitive imho. Especially for non-programmers who largely don’t use Git.
Not really, a branch is only a pointer to a commit, not to trees or blobs.
The commit in turn points to 1 or more parents (unless its the root-commit) and to the tree object that represents the state of the workspace at the time commit and contains pointers to blobs (files) and other trees (subdirectories).
For example, when I was first learning about git, I decided for some silly reason to make a git repo to store all of my little experiments—20 lines of python or matlab type stuff—just to have them all in one place. The main branch basically just has a README to remind me why the directory is empty, and to look at the other branches.
This is a very silly way to use git, I think, because they aren’t really branches of one coherent project. Diffing and merging them is useless. If they produce temporary output files, they all do it in totally non-standard ways that have no relation to one another. There really is no value to collecting them together. But it is possible, of course, because git can point to anything! Thankfully it is hiding on my hard-drive, hopefully nobody will ever see it other than me.
All of these phrases reinforce the common misconception that branches are “containers” for commits.
It would seem that you could query, “oh, what branch is this commit on?” But that’s not how git works. A commit has no idea what branch it is on, and that’s unintuitive. The only definition of being “on a branch” is, starting at the commit that the branch is currently pointing to, all ancestors of that commit are on the branch.
Having to walk the commit ancestry to see if a commit is on a branch doesn't make it not true.
Even despite knowing that's not how they work in git, it's usually easier for me to treat them that way.
A branch is a DAG of commits terminating at one commit.
They should have just called it a pointer if it's a pointer. But here we are anyway.
[Pedantically] this includes orphaned branches.
> They should have just called it a pointer if it's a pointer.
A branch in every VCS is first and foremost a name. That name has to evaluate to a) content, b) history. In Git and Git-like immutable-commit systems this can trivially be implemented by a table mapping branch (and tag and...) names to commits. I.e., in any immutable-commit VCS branches are naturally pointers.
The only real question is whether there should be additional semantics layered upon the pointer nature of branches. Indeed, there should be. In Git these are the basic rules:
- branch names are updated when adding commits in a workspace that has that branch checked out
- tag names are NOT updated when adding commits when adding commits in a workspace that has that tag checked out
- local configuration can enforce a variety of rules via commit hooks
- upstreams can enforce a variety of rules, such as fast-forward-only (no history rewriting)
This gives users a great deal of flexibility, including the ability to ban history rewriting on published branches and to allow it on unpublished branches.
The branch, like the tag, is a single reference to a commit.
Do this:
$ cat .git/refs/heads/mainThat said, it only makes sense to make such implicit mental shortcuts once you're already comfortable with git's data model. If you're not and you're still learning it, it's only going to confuse you.
The “dots on the line” visualization is pretty clear: you can have any number of branches pointing to a specific dot/commit. How do you add “branches” without fundamentally breaking that concept? When you create a branch, you’re not copying the line, you’re just adding a pointer.
(Also, note the path you ask us to cat does not contain the word "branch". You refer to how things are stored to support your definition, but the way things are stored doesn't use the term "branch".)
Plus that definition describes both tags and branches, how do you tell them apart? It’s clearly lacking. Actually that definition fits commits too.
I’d rather use these:
Branch: a pointer that advances
Tag: a permanent pointer
Commit/hash: an even more permanent pointer (because even a force can’t change it) that also describes the changes.
Because that is true; that is accurate. Thus, I don't understand your original claim that branches work in an unintuitive way.
To put this in programming terms, a git branch is a pointer. I think of a branch as the value is the most natural way to think about it.
This mismatch is what leads to endless tutorials on what git actually means when it uses its vocabulary.
The branch name is a pointer. All pointers (commits) are also branches of the DAG, just anonymous. Orphaned commits (i.e., not reachable from any named branch or tag) can be garbage collected (deleted).
The point is that raw pointer values do not make for user-friendly interfaces. Humans deal well with names and poorly with numbers. Thus we need names. Every VCS in the world supports branch names. The issue here is that Git's default rules for branches are shockingly simple for users coming from Mercurial and other VCSes, but those default rules are also the simplest possible default rules and you can add rules on top.
Like you, I have no problem working with the raw pointer values some of the time, so I often work in detached-HEAD mode, though in the end I have to always name commits I don't want to lose to GC and that I want to be able to find in the future -- in the end I still need names for commits, I still need named branches and tags.
One problem I'm trying to express with that question is: You say it's unintuitive, but when asked why you refer to a very low level technical detail.
My suspicion is that people find git branches unintuitive, and thus they think something more intuitive must exist. They have vague ideas about what "more intuitive" might mean, usually it means "just do what I want in this one situation I'm currently facing". My theory is that, if pressed, nobody can fully define a behavior for branches that is more intuitive than the current git implementation.
It shouldn't be controversial to say the porcelain can and should be much more straightforward to use.
DVCS concepts like the DAG and where you want to apply diffs are complicated the first time around, but after you get the concepts it should be easy to use, not difficult.
Mercurial's branches are a better match for how I've seen branches use in practice. Tags are essentially the same between both. Evolve's hidden commits make it easier to recover lost data than Git's reflog, though the underlying implementation not permitting garbage collection kind of sucks.
This comic presents a scenario where someone believes wrong information: git branches have rules. Git branches do not in fact have rules. This is wrong information. Why would we talk about wrong information?
We talk about wrong information because other people do not have up to date and correct information. We can try to just update their information to be correct, but that does not always work. In some cases they have built a structure of reasoning based on their belief. It is traumatic to destroy this structure of reasoning without providing a replacement.
Thus it can be helpful to understand the bad reasoning (based on wrong information) in order to communicate more effectively.
GitHub has ‘branch protection’ which prevents various changes. But GitHub is not Git. A common misconception lately due to its popularity.
Git itself is nothing more than a set of tools to manage the versions of a repository. How you choose to use it is up to you, and what you allow.
Solo developer? Rebase, or don’t.
Do you collaborate? Maybe consider using a SaaS product like GitHub and setup your rules.
"Git" is also the protocol, not just the implementation. Indeed, here I'm referring to the Git protocol more than to the Git implementation. There are multiple implementations of the Git protocol. GitHub is but one of several.
> But GitHub is not Git.
To the client it's indistinguishable.
Rebase/reset are major and first class features of Git. It's only wrong if you work with a team that expect the centralized server to be consistent.
Doing 'git pull' often would fit for a very specific workflow, and one that in time proves to be a wrong workflow, in my opinion.
Don't do things that don't have a conscious purpose.
In the linked comic she is pointing out, correctly, that while there are conventions that make good workflows, git doesn't enforce them and allows you to do stupid stuff. Surely you don't disagree with that.
$ git checkout -b test
Switched to a new branch 'test'
$ git checkout -b test/1
fatal: cannot lock ref 'refs/heads/test/1': 'refs/heads/test' exists; cannot create 'refs/heads/test/1'
Yep, it works exactly how you're assuming. You're asking the filesystem to create a directory with the same name as an existing file.
Por exemplo:
ls .git/refs/heads
cat .git/refs/heads/master % git init
Initialized empty Git repository in /private/tmp/gittest/.git/
% git commit --allow-empty -m "initial"
[main (root-commit) fa2cf4e] initial
% git checkout -b test
Switched to a new branch 'test'
% git checkout -b test/1
fatal: cannot lock ref 'refs/heads/test/1': 'refs/heads/test' exists; cannot create 'refs/heads/test/1'
% git --version
git version 2.39.3 (Apple Git-145)
% git checkout -b foo/bar
Switched to a new branch 'foo/bar'
% git checkout -b foo/qux
Switched to a new branch 'foo/qux'
% git checkout -b foo
fatal: cannot lock ref 'refs/heads/foo': 'refs/heads/foo/bar' exists; cannot create 'refs/heads/foo'
% pwd
/tmp/gittest/.git/refs/heads
% find . -print
.
./test
./foo
./foo/qux
./foo/bar
./main
% file *
foo: directory
main: ASCII text
test: ASCII text
You get even _more_ interesting problems if part of your team has case-insensitive file-systems!
Upstream branches can and do have rules, like "fast-forward only pushes", which means "no history rewriting here (because this branch is published)".
What you do in your own clone of the repository is between you and god alone.