Undo Git Rebase

Question

I performed a git rebase master on my branch and didn't realize it wasn't what I wanted until after I pushed it to the remote. It's only myself and 1 other person working on the project, so I know they've not pulled the latest changes.

Reading other questions on StackOverflow, it said to use git reflog and then git reset --hard HEAD@{n} to before the rebase. I did this to go to a commit I created before the rebase, but it didn't restore things to how it was before.

Am I missing a step? Is there a way to have the other person force-push his repo back up to restore things to how they were?

Thanks

Answer 1

As Makoto already noted, you probably shouldn't bother undoing this rebase: it's probably what you wanted. Nonetheless, feel free to read on for how to undo it.

---

Use the reflog for the branch, as it will be easier to read. (The HEAD reflog has the same information, but has a whole lot more stuff in it, hence it's harder to find what you're looking for.)

For instance if I had just rebased mybranch, I would see:

$ git reflog mybranch nnnnnnn mybranch@{0}: rebase finished: refs/heads/mybranch onto biguglysha1 ooooooo mybranch@{1}: commit: some sort of commit message ... 

The name mybranch@{1} is therefore synonymous (at the moment) with ooooooo, the old abbreviated SHA-1. Every time you do something to the branch (such as git reset) the number inside the @{...} part will change, while the SHA-1s are forever permanent, so it's a bit safer to use the SHA-1 (full or abbreviated) for cut-and-paste.

If you then:

$ git checkout mybranch # if needed 

and:

$ git reset --hard ooooooo  # or mybranch@{1} 

you should have the original back. This is because rebase simply copies commits and then moves the label. After the rebase, but before the reset, the commit graph looks something like this, where A through C are "your" commits:

          A - B - C              <-- (only in reflog now)         / ... - o - o - o - A' - B' - C'   <-- mybranch (after rebase) 

and git reset simply¹ erases the current branch label and pastes it on to the supplied SHA-1 (turning a reflog name into an SHA-1 first if needed). Hence, after the reset:

          A - B - C              <-- mybranch, plus older reflog         / ... - o - o - o - A' - B' - C'   <-- (only in reflog now) 

Note that now, post-reset, the rebase-made commit copies are the "abandoned" ones that are only found in reflog entries. The originals, which had been abandoned, are now claimed under mybranch again.

The way to think about this stuff is to draw the commit graph (with new commits pointing back at their parent commits), then draw in branch labels with long arrows pointing to the commit graph. The graph never² changes except to add new commits, which have new and different big ugly SHA-1s (which is why I use letters like A B and C instead, and tack on additives like A' for copies). The SHA-1s are guaranteed to be unique³ and are permanent, but the labels with their long arrows, get erased and re-pointed all the time. (If you do this on a whiteboard, you should generally use black for the commit graph and a color, or several colors, for the labels.)

---

¹Well, git reset does more than just move the label, unless you add some command-line flags. By default, it moves the label and resets the index; with --hard, it moves the label and resets the index and cleans out your work-tree. With --soft it just moves the label, leaving the index and work-tree alone. With git being what it is, there are a bunch more flags that twist up the meaning even further, but those are the big three: --soft, nothing aka --mixed, and --hard.

²If git only ever added things, your repo would grow huge over time. So, eventually, "unreachable" commits—those with no labels, and not even any leftover reflog entries, pointing to them, and that are not pointed-to by some commit that does have a label, or some other pointed-to commit—eventually these unreachable commits (and any other unreachable objects) are removed when git runs git gc for you automatically. You can force them to be removed earlier, but there's rarely a good reason to bother.

³Git itself depends on the guarantee. It's mathematically possible, but extremely improbable, for any two different objects to wind up with the same SHA-1. If this happens, git breaks.⁴ If the distribution is good enough the probability is 1 out of 2¹⁶⁰, which is really tiny. This is a good thing as the "Birthday paradox" raises the possibility pretty rapidly, but because it started so tiny, it stays tiny and in practice it's never been a problem.

⁴By design, the "breakage" is that git simply stops adding objects, so that everything-so-far is still good. You then have to move to whatever new system has been devised to handle billions-of-objects repositories, the "next generation" git or whatever.

Answer 2

I have good news and bad news for you.

The good news is, you're in the state that you express that you want! The last commit on master is indeed now in your branch, and all is well with the universe.

The bad news is that now, with your rebase operation, you have effectively moved your branch to the tip of master; that is, your development branch is now in the same state as if you had simply branched from the tip of master.

But this is [considered to be] a good thing: provided that you haven't had any merge conflicts, you haven't lost any data in this process, and you also don't have a worthless merge commit on your branch.

The book does a great job explaining what a rebase is, so I won't repeat it. From the operations you describe, you shouldn't have lost any work.

---

If you do want to recreate your original history before you force-pushed, then someone with a non-pulled history can force-push their branch up to the remote repository. This will cause the state of your repo to be whatever was on their box at the time they pushed.