After more than a decade of using git, I know that my comfort with rebase, and confidence that I wont make an error I can't undo, comes from knowing that I can always abort. And assuming it wasn't garbage collected, I can always get back to orphaned commits, or the commit before I rebased, as long as I keep their hashes. I can definitely look back on my less confident days as times when I wrongly assumed I was walking a tightrope where screwing up was painful and expensive, and easy to take the wrong path, and abort is the universal solvent to all of that.
I do still find myself tripping up on whether the next appropriate step is to make a commit or continue the rebase, as that is dependent on if you're in a merge conflict or just editing a commit. But even when I get that wrong, and collapse two commits together, abort saves the day.
This is a perfect use for tags. They allocate no extra storage space, and act as "savepoints" which you can refer to at any time. (Branches move, tags stay put.) They also guarantee that those loose ends are not garbage collected. Sometimes I delete a bunch of old ones, but any codebase of mine will at any time have a handful of old and probably useless tags. But that's ok.
tags and branches are both just refs to the commit and can be used interchangeably for something like this. tags are not immutable, just in convention.
You don't need to keep their hashes because git keeps them for you. If you realise too late that things went terribly wrong, you can get the pre-rebase hashes with "git reflog". (It can take a little getting used to knowing how to identify them quickly after a rebase but they're there.)
Better yet, git reset ORIG_HEAD or whatever is usually sufficient[0]. The reflog is the general solution, but git's magic references do provide quite a few niceties.
I use interactive rebases frequently, but even then, I think conflicts happening during a rebase are often somewhat cryptic.
I think I know what happens in principle (each step of a rebase does a merge of the commit from the list with the last rebased commit) but it's still often hard to understand which side of the conflict represents what, and what would be the desired outcome for that step in the history. So I tend to abort the rebase when I get a conflict and see if I can plan it differently so I can get it through without conflicts.
What I found good to know is that only actions that modify the actual changes in the history - reordering, deleting or editing commits - can cause conflicts. The other actions - reword, fixup or squash - only modify how the changes are grouped into commits and cannot cause conflicts.
So I usually do an interactive rebase in several passes: First a rebase that only reorders commits (or rarely does deletes or edits) and where I deal with the resulting conflicts, then a second pass for fixup or squash operations that can then use the reordered history from the first pass.
Rewords are both harmless and don't require any particular ordering, so can be done in any pass.
> but it's still often hard to understand which side of the conflict represents what, and what would be the desired outcome for that step in the history.
Are you using diff3/zdiff3? I ask because you seem to be describing exactly the problem it solves, or at least the way I try to sell people on it.
Basically in addition to 'current' & 'incoming from the rebase' hunks you get 'parent commit of incoming from the rebase' – which allows you to see 'what was I trying to change', i.e. how does the intent of it apply to the different thing that's now on master (whatever I'm reading onto).
I’ve used Git for 20 years now and wrote a book about it.
The biggest mistake most sources make when teaching about history rewriting is omitting these two lessons first:
- it’s easy to lose uncommitted data in Git and hard to lose committed data
- given this, learn how to use ‘git reflog’ to see what you’ve done and how to undo/redo it (as this post recommends but, even then, a little too late)
If you just get in the habit of constant committing and checking reflog: you’ll never lose data.
(Related lesson from my 10 years employed by GitHub: they almost never do the git gc you’d expect to remove commits so once you push a commit: it’s likely there forever and identifiable by that same hash from anywhere in the fork network. This can be bad/scary so be aware).
> (Related lesson from my 10 years employed by GitHub: they almost never do the git gc you’d expect to remove commits so once you push a commit: it’s likely there forever and identifiable by that same hash from anywhere in the fork network. This can be bad/scary so be aware).
The most important thing to happen to git in its entire history is jj as a better porcelain to git, so that you don't need to read entire books to get it right.
My fear has never been rebasing itself, but that I make some kind of mistake with a random change conflict and don't notice it because it's not tested for directly. But I'm also not sure if that's categorically a git fear. (Edit: related to this, I think this sort of minor "itchy worry" is grounded in how sometimes conceptually simple diffs look messy during conflict resolution.)
Like all git operations that destroy history, it's a bit less scary than rm or unlink, because you can undo obvious fuckups using the reflog. Non-obvious fuckups will (by Murphy's law) only be detected long after the reflog has been GC'd, so it still risks losing work.
Git was written with Linux in mind, where curating a clean commit history is more important than the code itself and warrants the extra hassle.
An actually usable rebase would probably require something like a meta history. But then the meta history would accumulate fixup "commits" and typos in commit messages so the OCD people will want to change that and we're back to square one.
In practice, the two times per decade it actually makes business sense to dig through the history to find when a bug was introduced instead of just writing the 5-line fix, you can live with a couple of imperfect commits and merges.
If all I'm doing with git rebase is reorganising my commits, and I have to deal with conflicts, after I'm done, I always git diff my pre-rebase head with post rebase head commit hash. I can quickly see I've messed something up if the diff isn't empty. For more complex tasks it's best to leave the rearranging until last, but where not possible, the diff is still pretty useful to double check my work.
made most of my fears regarding conflict resolution go away, as it right away gives me the stuff that I used to look for manually to understand the context of the conflict, which was probably the most error prone part of the process.
Reflog is, as the article says, the permission to do all kinds of crazy things, including rebasing. Once you realize any committed state is recoverable for a good period of time it really frees you.
Plus you can always take a note of the commit hash before you do anything crazy. But reflog means you don't have to.
Git really does have all the features you need. I wouldn't argue with anyone complaining about the UI to get there. But it's all in there somewhere.
This notion that `git rebase -i` is some scary unusual thing is totally alien to me. Doing some minor cleanup on local history with rebase is a *basic* operation that I would have thought everyone would be using regularly.
When I interview someone and they are scared of `git rebase -i` it's a huge red flag.
Git rebase has such foundational data structures and concepts that you couldn't pass a serious college programming course without grokking it.
The only reason someone would be scared of git rebase is:
a) they don't know the basics of programming/software engineering
b) they are not intellectually curious enough to look into how tf the software they use every day works
You probably shouldn't have engineers like that on your team.
Understanding git rebase is important. You aren't expected to rebase mainline branches regularly but if your team operates on a merge/patchset workflow and isn't just squashing everything down onto main each PR you need to understand how to rebase your changes.
You do your development and use rebase to organise your code into logical commits/patches that can each be reviewed separately. This way when you do your review in github or on the mailing list or whatever you can review each logical change by going through and reviewing the commits separately (which you can do in github by selecting the commit in the dropdown for the review window).
And of course via email based workflows or "stacked patchset" workflows like what tangled provides then you can get deltas between revisions so that you can see what changed in a given commit between rev 1 and rev 2 so that you don't have to re-review the entire patchset/PR, only the stuff that matters (but still broken up into logical changes).
I wonder if it's because it opens in the default visual editor (typically vi/vim) and while I'm primarily a (neo)vim user now, it is certainly not the most intuitive editor to get started with without guidance.
The part where git needs to be “understood” is the entire problem. “Do one thing and do it well” was the whole mantra, which was completely ignored with the disaster that is git. It’s objectively awful.
You only have to learn three things to use a computer: Your shell, your editor, and your version control system.
That's it. That basic knowledge will carry you far, and will be useful a decade from now. Those are the three best afternoons you can spend. Do it out of respect for the computer.
What's objectively awful is your incredibly intellectually lazy take and how widespread it is. It's always worth taking a few hours to understand powerful tools that you're going to spend years/decades with, and Git isn't even that hard to grasp when you actually want to learn rather than make excuses for your unwillingness to spend a bit of effort on it.
Over time I'm becoming more and more convinced that the only major Git's sin (and the reason for its popularity at the same time) is being so easy to use that it can be used for basic things with no understanding whatsoever, so people just cargocult it and call it a day. Then once they accidentally get off the beaten path they are absolutely hopeless against the most trivial issues they're being presented with.
Git is a tool for manipulating a data structure that is the repository. Using it without understanding that structure is like trying to use a word processor without knowing how to write - and apparently half of the industry is content with only knowing how to put clipart onto the printed page.
It could be so much better. At the same time, it’s one of the core tools for a developer. You’ll learn it eventually. You’re going to be using version control for decades. I’ve met devs many years into their careers who don’t understand rebase. It’s like going to a metal shop and they’re complaining about the learning curve on their welder and half the tradesmen are still at a novice level.
I still don't know it very well, and it's been well over a decade. I think it's a combination of a few things:
- I find it uninteresting. My version control needs are very simple.
- Most teams I have been in use a small subset of it.
- It's confusing terms and inconsistent cli are huge warning signs to not go down that rabbithole. Today instead of learning Git I read some Tony Hoare, much better.
I've been at once place where they rebased, and it was awful. Complete waste of time. (great place otherwise though)
One additional thing I would mention ist that: When resolving conflicts never try to solve them for the final result. Consider each conflict without considering how the code will change in a later commit. I’ve seen people die a painful rebase death because of this.
Yeah like the other commenter below mentioned, `rerere` takes a lot of the pain out of layering rebased changes. It's important (IMO) to make sure each commit still contains a logical change and a working system, even with an incremental rebase.
Why isn't the 'edit' command mentioned? It's one of the most useful features. It can be used for splitting commits, for example. Mark the commit you want to split to be edited. The commit replay is going to stop after that commit. Now:
git reset HEAD^1 (NO --hard!)
git commit --patch ...
git rebase --continue
You want to add some pending changes to the previous commit? No problem:
git commit --patch --amend ...
You want to move the pending changes to future commits?
I wish my team would at least use the rebase merge strategy by default to avoid branch graphs looking like the London tube network. Not in 10 years did I have an issue with that (okay, except accidentally rebasing deliberate no-ff merges).
The one time I suggested that, someone immediately came up to me trying to convince me that rebases are the most dangerous thing ever.
I worked on a team that did this, + required feature branches to get squashed to a single commit before merge. On that commit, we'd require a very brief bullet list of changes, sometimes a POC, and a link to the PR. This made git blame a lot more helpful when debugging issues. I am a big fan of this approach.
> On that commit, we'd require a very brief bullet list of changes, sometimes a POC, and a link to the PR.
one of the first things i do in a new gitlab repo is set up ff+squash commit merges with the squash commit message template automatically pulling the MR title, link, description, authors etc.
I found the VS Code GitLens extension to be a good abstraction for interactive rebase. It provides a drag-and-drop UI with a dropdown to select actions applied to each commit. That is much easier than editing a text file.
Oh that's really very clean, yeah. It's almost literally exactly what's in the text editor, just with a more friendly-for-mouse UX. They did very well.
I was hoping for a rebase that recalled what your previous base commit of your branch.
I'm really not fond of having to `git rebase --interactive` just to drop all commits that are "already in master", especially when I forget this silly step and get a borked rebase.
I use that one very frequently (after a fixup, of course) because I want the author date on amended commits to reflect the last time I edited them, not the first time I committed them.
If you do a rebase -i because you want to squash or fix some commits, then there's git commit --squash and git commit --fixup commands which will mark commits as intending to squash or fixup another commit, and then you can use git rebase --autosquash <base> to automatically do what you want in one go. Of course, often you forget to use the --squash/--fixup options or you can't be bothered to lookup the hash of the commit you want to fix, so you'll end up using git rebase -i anyway.
I still make a throwaway branch before a complicated rebase. It costs nothing and turns the whole operation into a low-stakes edit instead of a leap of faith.
Git rebase isn't scary at all, what's scary is the git rebase cult. I'll probably die never really seeing the tangible benefit a meticulously curated git history provides but at least I saw a few hundred blog posts and hn comments declaring the moral failure that is not having the cleanest commit history. The cynic in me would say that people are over-identifying too much with their recent bikeshedding addiction (just like 10-page vim configs, custom built keyboards or whatever else promises them immense but immeasurable performance gains) but in reality I guess it's just completely different mental models about code. I have yet to look at any code or it's however dirty history and needed to consult a commit message to understand what it's there for but I'm looking forward to the day where one of my rebase colleagues fixes a bug with git bisect that wasn't efficiently fixable any other way.
People are different, and I think we have to respect that. In my role I tend to have to fix problems in other people's code probably on a weekly basis. Commits that do too much, or too little, or are generally hard to understand are the number one reason this takes time. Crafting readable commits takes seconds out of your work day. Please be considerate to all your future colleagues.
I use git history to understand wtf my colleagues, and myself, where doing and thinking 10 years ago. A well structured, squashed commit and a clean history helps A LOT. I can't tell you how often I find a 'Apply review changes' and 'Fix shit' commit with no context. I then need to go back in the blame history or find what branch/PR this belonged too on GH to find even the context of these changes.
Perhaps you don't work on the kind off crappy code bases I am dealing with, but a bit of git discipline saves a fuckton of frustrations.
tbf, this says a lot more about the code base, my colleagues and the shitty, lazy culture at our company that leads to ZERO ownership and architecture ;)
Nice balance of practical examples and mental models. One question though, how do you decide when to use interactive rebase versus relying on squash merges in platforms like GitHub.
Adopting interactive rebasing along with curating my commits for the benefit of the reviewer has been the best upgrade I've made to my git usage in probably the last five years.
Most developers just have a linear view of git, which is fine. Should I really care that origins are just named nodes on combined graphs? Like, I just want to check in my code before I have to merge someone else's.
I should publish it at some point (it's just a simple bash script), but I made a convenient alias `git bisect-rebase` which helps catch up very old branches.
It isn't particularly fast, but it uses bisect to find the first commit on the target branch that conflicts with your branch, then let's you rebase to that commit. You then repeat the process until you are caught up.
The idea is that solving one conflict many times is usually quite easy, especially if you know the conflicting commit messages, but solving many conflicts in one go is not so easy. This was inspired by me rebasing a 6 month old branch which refactored a file that happened to have many edits in between.
please people, stop teaching newcomers short aliases. please teach them proper long terms --interactive is not much harder to type than -i. Why do we keep doing this? It isn't the 1970s anymore. even if we do this all day it will take what maybe 5 KiB more of ram or storage? if that? why teach people some Harry Potter incantations when we can teach them proper words?
Sure, if they want to they can later use `npm -g i` or whatever abomination they want but when we teach them, we should use full arguments, not aliases. please, people :(
Rebase is scary if you are conflicting on many items.
I have a simple rule where if the rebase cannot be solved within a few commit rewrites that I will restart the branch from latest master.
If you are suffering from really difficult rebase scenarios, it's likely that the senior developers are more at fault than the junior developers. The way you organize work over the codebase has the biggest impact on how things would conflict. If nothing ever does, rebasing is a trivial operation.
There's nothing scary about conflicts on cherry-picked commits. It's just sometimes tedious to resolve them one-by-one, and sometimes may just not be worth the effort if it can be done in a better way.
Another git command I love using is `git add -p`. I always want to commit in chunks as I go but sometimes I wait to long or realize later I could have broken it up. Being able to select only pieces of a file to add to a specific commit is so great. Editing the chunks manually can be a bit intimidating but since it only stages the change you can always restore the file and try again if the diff doesn't look right after the add.
That "git add -p" works from the command line is cool, but this is the one git task that I feel is significantly easier using a GUI tool (I use one specifically for this (and viewing diffs) while otherwise mostly using the command line).
I used to feel the same, and I still use a TUI tool, but after working with people who always used the CLI for this (who oddly otherwise were big into IDEs) it's actually an oddly really nice experience and not really any faster or slower. In fact, seeing a hunk at a time has this "focusing" affect on thinking about each change.
> it is very hard to actually lose work here, for three reasons.
The opposite is true, of course, for example: you move commits arround, start to resolve conflicts, and after a few steps you realize you can't, so when you
> you can bail out at any time. as aforementioned with git rebase --abort
Yep, except for all the work you've done resolving the conflict, that's "aborted"
> the old commits still exist in git’s object database, unreferenced but intact,
The best UI there is - some hidded data store you're, of course, intimately familiar with, so won't have any trouble getting data out of. Oh, wait, you didn't even know it existed? Tough luck, git gud to avoid data loss!
> botched rebase is a few minutes of you perusing through the reflog.
Unless, of course, your not that sharp to remember all the commits from their names and would also like to see the diff contents to connect to your code work. But that's just more minutes extra, not that big of a loss.
> there’s of course, the low-tech insurance policy: git branch backup-before-rebase
Finally some sensible advice, one that should be the default backup plan to save userst he trouble of wasting "few minutes" perusing the logs
pick a1b2c3d Add user model
s e4f5g6h Fix typo in user model
pick i7j8k9l Add login endpoint
s m0n1o2p WIP debugging login endpoint
s feedbee fixed login endpoint
s deadbee fixed login endpoint
s adebade fixed login endpoint
s abcdefg actually fixed login endpoint
Of course, I get reminded about f/fixup every now and then - think "golly, that will save a second or two" and promptly forget.
I typically work with commit chains of at least 5 commits in various states of code review, so I spend at least 60% of my development time editing an old commit in the middle of a rebase -i.
sometimes i wish mercurial won, if nothing else for the fact that i found its ux to be more enjoyable. I respect and use git, but never got comfortable with merging. im thankful AI can automate it for me now
The nice thing with magit is how easy it is to manipulate hunks and line, truly editing commits with a surgical precision, all with an handful of keybindings.
rebase interactive is so great, such a powerful tool.
it's the one clear tool that i used all the time. moving to jj, it has lots of amazing tools starting with `jj edit` to go change the past safely & conveniently, with much less pain. but i miss the direct powerful data-driven experience of seeing the timeline in a text file, and picking what to do with it. sometimes.
imo every dev should work to get some competency with git rebase! it's an amazing tool. there's few other systems that give such direct control. it's top layer is, perhaps, the excel of version control?
After more than a decade of using git, I know that my comfort with rebase, and confidence that I wont make an error I can't undo, comes from knowing that I can always abort. And assuming it wasn't garbage collected, I can always get back to orphaned commits, or the commit before I rebased, as long as I keep their hashes. I can definitely look back on my less confident days as times when I wrongly assumed I was walking a tightrope where screwing up was painful and expensive, and easy to take the wrong path, and abort is the universal solvent to all of that.
I do still find myself tripping up on whether the next appropriate step is to make a commit or continue the rebase, as that is dependent on if you're in a merge conflict or just editing a commit. But even when I get that wrong, and collapse two commits together, abort saves the day.
This is a perfect use for tags. They allocate no extra storage space, and act as "savepoints" which you can refer to at any time. (Branches move, tags stay put.) They also guarantee that those loose ends are not garbage collected. Sometimes I delete a bunch of old ones, but any codebase of mine will at any time have a handful of old and probably useless tags. But that's ok.
tags and branches are both just refs to the commit and can be used interchangeably for something like this. tags are not immutable, just in convention.
You don't need to keep their hashes because git keeps them for you. If you realise too late that things went terribly wrong, you can get the pre-rebase hashes with "git reflog". (It can take a little getting used to knowing how to identify them quickly after a rebase but they're there.)
Better yet, git reset ORIG_HEAD or whatever is usually sufficient[0]. The reflog is the general solution, but git's magic references do provide quite a few niceties.
https://www.man7.org/linux/man-pages/man7/gitrevisions.7.htm...
> as long as I keep their hashes
that means my hashes or their hashes or them keeping my hashes, I am never sure
you can just make a backup copy of .git or your entire project dir. When you get lost, just delete your .git, copy from the backup and try again
You never need to do that. git-reflog(1) is the most general solution if you will ever need.
Is this a joke? git's whole purpose is to save the history of your project and return to earlier versions when needed
It's a super simple solution for which you can be certain that you have all the knowledge needed to perform the restore operation.
I use interactive rebases frequently, but even then, I think conflicts happening during a rebase are often somewhat cryptic.
I think I know what happens in principle (each step of a rebase does a merge of the commit from the list with the last rebased commit) but it's still often hard to understand which side of the conflict represents what, and what would be the desired outcome for that step in the history. So I tend to abort the rebase when I get a conflict and see if I can plan it differently so I can get it through without conflicts.
What I found good to know is that only actions that modify the actual changes in the history - reordering, deleting or editing commits - can cause conflicts. The other actions - reword, fixup or squash - only modify how the changes are grouped into commits and cannot cause conflicts.
So I usually do an interactive rebase in several passes: First a rebase that only reorders commits (or rarely does deletes or edits) and where I deal with the resulting conflicts, then a second pass for fixup or squash operations that can then use the reordered history from the first pass.
Rewords are both harmless and don't require any particular ordering, so can be done in any pass.
> but it's still often hard to understand which side of the conflict represents what, and what would be the desired outcome for that step in the history.
Are you using diff3/zdiff3? I ask because you seem to be describing exactly the problem it solves, or at least the way I try to sell people on it.
Basically in addition to 'current' & 'incoming from the rebase' hunks you get 'parent commit of incoming from the rebase' – which allows you to see 'what was I trying to change', i.e. how does the intent of it apply to the different thing that's now on master (whatever I'm reading onto).
Thanks for the info! I didn't try that so far, but it sounds like it could really make that easier. Will give it a try!
In case of many commits with also many complex merge conflicts, I often find it easier to just:
1. Abort rebase.
2. Squash all my commits into a single commit.
3. Rebase on target branch again.
I’ve used Git for 20 years now and wrote a book about it.
The biggest mistake most sources make when teaching about history rewriting is omitting these two lessons first:
- it’s easy to lose uncommitted data in Git and hard to lose committed data
- given this, learn how to use ‘git reflog’ to see what you’ve done and how to undo/redo it (as this post recommends but, even then, a little too late)
If you just get in the habit of constant committing and checking reflog: you’ll never lose data.
(Related lesson from my 10 years employed by GitHub: they almost never do the git gc you’d expect to remove commits so once you push a commit: it’s likely there forever and identifiable by that same hash from anywhere in the fork network. This can be bad/scary so be aware).
> (Related lesson from my 10 years employed by GitHub: they almost never do the git gc you’d expect to remove commits so once you push a commit: it’s likely there forever and identifiable by that same hash from anywhere in the fork network. This can be bad/scary so be aware).
Oh wow. Is there a way to trigger GC manually?
For github itself the answer is "contact them if you think you pushed a secret".
For git. You can run `git maintenance run` or the legacy command `git gc`.
Unsure, no longer there. I think maybe just “contact support”.
The most important thing to happen to git in its entire history is jj as a better porcelain to git, so that you don't need to read entire books to get it right.
> it’s likely there forever and identifiable by that same hash from anywhere in the fork network
in case you are not aware of it already:
https://www.aikido.dev/blog/the-fork-awakens-why-githubs-inv...
After two decades of using git I'm either telling LLMs to do rebases for me or use jj (or both). Life's too short.
[delayed]
I feel like if you're scared of rebasing, you don't actually understand git.
My fear has never been rebasing itself, but that I make some kind of mistake with a random change conflict and don't notice it because it's not tested for directly. But I'm also not sure if that's categorically a git fear. (Edit: related to this, I think this sort of minor "itchy worry" is grounded in how sometimes conceptually simple diffs look messy during conflict resolution.)
Like all git operations that destroy history, it's a bit less scary than rm or unlink, because you can undo obvious fuckups using the reflog. Non-obvious fuckups will (by Murphy's law) only be detected long after the reflog has been GC'd, so it still risks losing work.
Git was written with Linux in mind, where curating a clean commit history is more important than the code itself and warrants the extra hassle.
An actually usable rebase would probably require something like a meta history. But then the meta history would accumulate fixup "commits" and typos in commit messages so the OCD people will want to change that and we're back to square one. In practice, the two times per decade it actually makes business sense to dig through the history to find when a bug was introduced instead of just writing the 5-line fix, you can live with a couple of imperfect commits and merges.
If all I'm doing with git rebase is reorganising my commits, and I have to deal with conflicts, after I'm done, I always git diff my pre-rebase head with post rebase head commit hash. I can quickly see I've messed something up if the diff isn't empty. For more complex tasks it's best to leave the rearranging until last, but where not possible, the diff is still pretty useful to double check my work.
range-diff can be useful for that too.
Setting:
made most of my fears regarding conflict resolution go away, as it right away gives me the stuff that I used to look for manually to understand the context of the conflict, which was probably the most error prone part of the process.Reflog is, as the article says, the permission to do all kinds of crazy things, including rebasing. Once you realize any committed state is recoverable for a good period of time it really frees you.
Plus you can always take a note of the commit hash before you do anything crazy. But reflog means you don't have to.
Git really does have all the features you need. I wouldn't argue with anyone complaining about the UI to get there. But it's all in there somewhere.
This notion that `git rebase -i` is some scary unusual thing is totally alien to me. Doing some minor cleanup on local history with rebase is a *basic* operation that I would have thought everyone would be using regularly.
When I interview someone and they are scared of `git rebase -i` it's a huge red flag.
Git rebase has such foundational data structures and concepts that you couldn't pass a serious college programming course without grokking it.
The only reason someone would be scared of git rebase is:
You probably shouldn't have engineers like that on your team.An interviewer asking me about git rebase is the huge red flag
Understanding git rebase is important. You aren't expected to rebase mainline branches regularly but if your team operates on a merge/patchset workflow and isn't just squashing everything down onto main each PR you need to understand how to rebase your changes.
You do your development and use rebase to organise your code into logical commits/patches that can each be reviewed separately. This way when you do your review in github or on the mailing list or whatever you can review each logical change by going through and reviewing the commits separately (which you can do in github by selecting the commit in the dropdown for the review window).
And of course via email based workflows or "stacked patchset" workflows like what tangled provides then you can get deltas between revisions so that you can see what changed in a given commit between rev 1 and rev 2 so that you don't have to re-review the entire patchset/PR, only the stuff that matters (but still broken up into logical changes).
Genuine question, why? It seems like a reasonable knowledge check that can be answered in a few sentences.
I wonder if it's because it opens in the default visual editor (typically vi/vim) and while I'm primarily a (neo)vim user now, it is certainly not the most intuitive editor to get started with without guidance.
I'm genuinely unsure why rebase interactive is more scary than rebase
Yeah I never use non-interactive rebase. Even if I don't need to edit the todo list, reading it gives me a confirmation of what git will do.
Correct on both counts.
The part where git needs to be “understood” is the entire problem. “Do one thing and do it well” was the whole mantra, which was completely ignored with the disaster that is git. It’s objectively awful.
Absolutely everything in computing needs to be 'understood' to be used.
You only have to learn three things to use a computer: Your shell, your editor, and your version control system.
That's it. That basic knowledge will carry you far, and will be useful a decade from now. Those are the three best afternoons you can spend. Do it out of respect for the computer.
What's objectively awful is your incredibly intellectually lazy take and how widespread it is. It's always worth taking a few hours to understand powerful tools that you're going to spend years/decades with, and Git isn't even that hard to grasp when you actually want to learn rather than make excuses for your unwillingness to spend a bit of effort on it.
Over time I'm becoming more and more convinced that the only major Git's sin (and the reason for its popularity at the same time) is being so easy to use that it can be used for basic things with no understanding whatsoever, so people just cargocult it and call it a day. Then once they accidentally get off the beaten path they are absolutely hopeless against the most trivial issues they're being presented with.
Git is a tool for manipulating a data structure that is the repository. Using it without understanding that structure is like trying to use a word processor without knowing how to write - and apparently half of the industry is content with only knowing how to put clipart onto the printed page.
sounds like someone accidentally delete your repo because they didn't understand git and did a dumdum :P
It could be so much better. At the same time, it’s one of the core tools for a developer. You’ll learn it eventually. You’re going to be using version control for decades. I’ve met devs many years into their careers who don’t understand rebase. It’s like going to a metal shop and they’re complaining about the learning curve on their welder and half the tradesmen are still at a novice level.
I still don't know it very well, and it's been well over a decade. I think it's a combination of a few things:
- I find it uninteresting. My version control needs are very simple.
- Most teams I have been in use a small subset of it.
- It's confusing terms and inconsistent cli are huge warning signs to not go down that rabbithole. Today instead of learning Git I read some Tony Hoare, much better.
I've been at once place where they rebased, and it was awful. Complete waste of time. (great place otherwise though)
Have you ever read any of the introductory material that the git project itself maintains for teaching how to use the tool?
- https://git-scm.com/docs/gittutorial
- https://git-scm.com/docs/giteveryday
- https://git-scm.com/docs/gitworkflows
- https://git-scm.com/docs/gitfaq
- https://git-scm.com/cheat-sheet
Or if you want to sit down and really learn the nuts and bolts
- https://git-scm.com/book/en/v2
Give it an hour or two. It is really not that complicated. A little bit of investment will pay huge dividends. Do yourself the favor.
I could say the same for reading Tony Hoare.
yeah I would really love to see a subset of git with a sane cli. would switch in a heartbeat.
Git has been re-designing their tooling over the last few years. Nowadays there exist really pleasant, easy to use versions of the main tools.
switch and restore replaced checkout. history is being developed to replace rebase for most common use cases. maintenance replaces gc. etc.
Fossil?
The HN title is "Git rebase -I etc" but the post title is "Git rebase -i"
I guess a keyboard auto corrected i into I.
One additional thing I would mention ist that: When resolving conflicts never try to solve them for the final result. Consider each conflict without considering how the code will change in a later commit. I’ve seen people die a painful rebase death because of this.
Yeah like the other commenter below mentioned, `rerere` takes a lot of the pain out of layering rebased changes. It's important (IMO) to make sure each commit still contains a logical change and a working system, even with an incremental rebase.
rerere
this is not a cat on a keyboard. it is the name of a git command that does this.
Why isn't the 'edit' command mentioned? It's one of the most useful features. It can be used for splitting commits, for example. Mark the commit you want to split to be edited. The commit replay is going to stop after that commit. Now:
You want to add some pending changes to the previous commit? No problem: You want to move the pending changes to future commits?I found the following config to bring a great QoL improvement:
Now you can do quick commits as save-points, and squash all of them at the end.A workflow I was introduced to and have worked with a lot in my teams is,
1. Everyone uses feature branches
2. Everyone cleans up their branch using interactive rebase on top of main before review
3. All merges have been freshly rebase on main
I think it works very well and keeps the history clean while preserving "each commit does one thing".
I wish my team would at least use the rebase merge strategy by default to avoid branch graphs looking like the London tube network. Not in 10 years did I have an issue with that (okay, except accidentally rebasing deliberate no-ff merges).
The one time I suggested that, someone immediately came up to me trying to convince me that rebases are the most dangerous thing ever.
I worked on a team that did this, + required feature branches to get squashed to a single commit before merge. On that commit, we'd require a very brief bullet list of changes, sometimes a POC, and a link to the PR. This made git blame a lot more helpful when debugging issues. I am a big fan of this approach.
> On that commit, we'd require a very brief bullet list of changes, sometimes a POC, and a link to the PR.
one of the first things i do in a new gitlab repo is set up ff+squash commit merges with the squash commit message template automatically pulling the MR title, link, description, authors etc.
I found the VS Code GitLens extension to be a good abstraction for interactive rebase. It provides a drag-and-drop UI with a dropdown to select actions applied to each commit. That is much easier than editing a text file.
Here's a GIF I found with Google: https://yogwang.site/2025/cursor-vscode-gitlens-rebase-edito...
Oh that's really very clean, yeah. It's almost literally exactly what's in the text editor, just with a more friendly-for-mouse UX. They did very well.
That “-I” really needs to be lowercase.
And, ironically, a lot of letters in the blog post need to be uppercase.
Someone should collect all the cases when header capitalization on HN affected meaning
Yeah I got quite excited that there was somehow a better -i; its almost perfect clickbait
I was hoping for a rebase that recalled what your previous base commit of your branch.
I'm really not fond of having to `git rebase --interactive` just to drop all commits that are "already in master", especially when I forget this silly step and get a borked rebase.
Does this not do what you want?
The gitrevisions(7) manpage is a good reference to keep in your back pocket, I find.https://www.man7.org/linux/man-pages/man7/gitrevisions.7.htm...
I wish there were a shorthand for
I use that one very frequently (after a fixup, of course) because I want the author date on amended commits to reflect the last time I edited them, not the first time I committed them.But you can make your own alias?
https://git-scm.com/book/ms/v2/Git-Basics-Git-Aliasesi edit my commit dates constantly for various reasons
get ahead at work? split into commits and make it look like multiple days of work
work on your own IP during work hours? edit the timestamp to be outside work hours (i even have a script for this)
If you do a rebase -i because you want to squash or fix some commits, then there's git commit --squash and git commit --fixup commands which will mark commits as intending to squash or fixup another commit, and then you can use git rebase --autosquash <base> to automatically do what you want in one go. Of course, often you forget to use the --squash/--fixup options or you can't be bothered to lookup the hash of the commit you want to fix, so you'll end up using git rebase -i anyway.
I still make a throwaway branch before a complicated rebase. It costs nothing and turns the whole operation into a low-stakes edit instead of a leap of faith.
Yes, and you can diff the two branches when you are done!
Git rebase isn't scary at all, what's scary is the git rebase cult. I'll probably die never really seeing the tangible benefit a meticulously curated git history provides but at least I saw a few hundred blog posts and hn comments declaring the moral failure that is not having the cleanest commit history. The cynic in me would say that people are over-identifying too much with their recent bikeshedding addiction (just like 10-page vim configs, custom built keyboards or whatever else promises them immense but immeasurable performance gains) but in reality I guess it's just completely different mental models about code. I have yet to look at any code or it's however dirty history and needed to consult a commit message to understand what it's there for but I'm looking forward to the day where one of my rebase colleagues fixes a bug with git bisect that wasn't efficiently fixable any other way.
People are different, and I think we have to respect that. In my role I tend to have to fix problems in other people's code probably on a weekly basis. Commits that do too much, or too little, or are generally hard to understand are the number one reason this takes time. Crafting readable commits takes seconds out of your work day. Please be considerate to all your future colleagues.
I use git history to understand wtf my colleagues, and myself, where doing and thinking 10 years ago. A well structured, squashed commit and a clean history helps A LOT. I can't tell you how often I find a 'Apply review changes' and 'Fix shit' commit with no context. I then need to go back in the blame history or find what branch/PR this belonged too on GH to find even the context of these changes. Perhaps you don't work on the kind off crappy code bases I am dealing with, but a bit of git discipline saves a fuckton of frustrations.
tbf, this says a lot more about the code base, my colleagues and the shitty, lazy culture at our company that leads to ZERO ownership and architecture ;)
I continual find the amount of ink spilled on having a clean commit history amazing, when you can basically get all the same benefits with:
git log --first-parent
git bisect --first-parent
I fix bugs will git bisect all the time. But it’s not practical if you don’t have a good commit history to begin with.
Nice balance of practical examples and mental models. One question though, how do you decide when to use interactive rebase versus relying on squash merges in platforms like GitHub.
You dont need to open the reflog you can just
After a rebase to "undo" it.Adopting interactive rebasing along with curating my commits for the benefit of the reviewer has been the best upgrade I've made to my git usage in probably the last five years.
A young dev who's still ambitious, gotta love it.
Most developers just have a linear view of git, which is fine. Should I really care that origins are just named nodes on combined graphs? Like, I just want to check in my code before I have to merge someone else's.
Zero rebases since Fable 5.
I should publish it at some point (it's just a simple bash script), but I made a convenient alias `git bisect-rebase` which helps catch up very old branches.
It isn't particularly fast, but it uses bisect to find the first commit on the target branch that conflicts with your branch, then let's you rebase to that commit. You then repeat the process until you are caught up.
The idea is that solving one conflict many times is usually quite easy, especially if you know the conflicting commit messages, but solving many conflicts in one go is not so easy. This was inspired by me rebasing a 6 month old branch which refactored a file that happened to have many edits in between.
That somewhat sounds like the sort of thing git imerge does.
please people, stop teaching newcomers short aliases. please teach them proper long terms --interactive is not much harder to type than -i. Why do we keep doing this? It isn't the 1970s anymore. even if we do this all day it will take what maybe 5 KiB more of ram or storage? if that? why teach people some Harry Potter incantations when we can teach them proper words?
Sure, if they want to they can later use `npm -g i` or whatever abomination they want but when we teach them, we should use full arguments, not aliases. please, people :(
Rebase is scary if you are conflicting on many items.
I have a simple rule where if the rebase cannot be solved within a few commit rewrites that I will restart the branch from latest master.
If you are suffering from really difficult rebase scenarios, it's likely that the senior developers are more at fault than the junior developers. The way you organize work over the codebase has the biggest impact on how things would conflict. If nothing ever does, rebasing is a trivial operation.
There's nothing scary about conflicts on cherry-picked commits. It's just sometimes tedious to resolve them one-by-one, and sometimes may just not be worth the effort if it can be done in a better way.
I used to always have deep dread when having to use git but then I found [0]
[0] https://ohshitgit.com/
- i often go back like 5 commits and make changes like this
- git rebase -i <commit-hash>^
- select edit
- git reset --soft HEAD~
- make some changes
- git add . && git commit -m "changes"
- git rebase --continue
- Am I using git rebase wrong?
I wouldn't say it's wrong.
My preference, however, is creating a fixup commit and using rebase to squash it into the old one.
Edit: A sibling comment mentioned "git history fixup" which I'll have to try out.
No, but you would find jj solves this problem far more elegantly.
Another git command I love using is `git add -p`. I always want to commit in chunks as I go but sometimes I wait to long or realize later I could have broken it up. Being able to select only pieces of a file to add to a specific commit is so great. Editing the chunks manually can be a bit intimidating but since it only stages the change you can always restore the file and try again if the diff doesn't look right after the add.
That "git add -p" works from the command line is cool, but this is the one git task that I feel is significantly easier using a GUI tool (I use one specifically for this (and viewing diffs) while otherwise mostly using the command line).
I used to feel the same, and I still use a TUI tool, but after working with people who always used the CLI for this (who oddly otherwise were big into IDEs) it's actually an oddly really nice experience and not really any faster or slower. In fact, seeing a hunk at a time has this "focusing" affect on thinking about each change.
Again, though, I still use a visual tool.
Yeah, that's what I use the built-in `git gui` for.
Then you will really like:
https://git-scm.com/docs/git-history#Documentation/git-histo...> it is very hard to actually lose work here, for three reasons.
The opposite is true, of course, for example: you move commits arround, start to resolve conflicts, and after a few steps you realize you can't, so when you
> you can bail out at any time. as aforementioned with git rebase --abort
Yep, except for all the work you've done resolving the conflict, that's "aborted"
> the old commits still exist in git’s object database, unreferenced but intact,
The best UI there is - some hidded data store you're, of course, intimately familiar with, so won't have any trouble getting data out of. Oh, wait, you didn't even know it existed? Tough luck, git gud to avoid data loss!
> botched rebase is a few minutes of you perusing through the reflog.
Unless, of course, your not that sharp to remember all the commits from their names and would also like to see the diff contents to connect to your code work. But that's just more minutes extra, not that big of a loss.
> there’s of course, the low-tech insurance policy: git branch backup-before-rebase
Finally some sensible advice, one that should be the default backup plan to save userst he trouble of wasting "few minutes" perusing the logs
I also thought rebase was scary long ago. It is scary in the sense if one messes something up without realizing and then pushes to master…
Besides interactive rebase, my other favorite command is “git log —-oneline origin/master..HEAD”
It shows me exactly where I am…
Big rebase fan here. I find myself using squash the most.
I use rebase so much but it only feels comfortable because I follow a workflow while rebasing.
I typically work with commit chains of at least 5 commits in various states of code review, so I spend at least 60% of my development time editing an old commit in the middle of a rebase -i.
You might find `jj` to help you there: all changesets have stable identifier and their content can be updated at any moment in-place.
TLDR; `jj edit <what_to_edit>; $EDITOR` will update a commit in the middle of the branch
sometimes i wish mercurial won, if nothing else for the fact that i found its ux to be more enjoyable. I respect and use git, but never got comfortable with merging. im thankful AI can automate it for me now
I really love the emacs interactive rebase mode, which comes up by default when you have EDITOR=emacs and do rebase -i.
I know you can do this from within magit as well but I've never gotten into that workflow.
The nice thing with magit is how easy it is to manipulate hunks and line, truly editing commits with a surgical precision, all with an handful of keybindings.
Even just rebase fixup a commit that is a few commits below HEAD is just a few chords away. Use it many times a week when working.
rebase interactive is so great, such a powerful tool.
it's the one clear tool that i used all the time. moving to jj, it has lots of amazing tools starting with `jj edit` to go change the past safely & conveniently, with much less pain. but i miss the direct powerful data-driven experience of seeing the timeline in a text file, and picking what to do with it. sometimes.
imo every dev should work to get some competency with git rebase! it's an amazing tool. there's few other systems that give such direct control. it's top layer is, perhaps, the excel of version control?
I cannot recall the last time I committed manually, let alone rebase manually. An agent can do it for you faster and without making mistakes.