pull --rebase [LWN.net]

pull --rebase

Posted Mar 20, 2012 0:46 UTC (Tue) by dlang (✭ supporter ✭, #313) [Link]

rebasing means that you end up with a new version of code that you have never tested, and it's impossible for anyone looking at the resulting tree to figure out what you developed against.

The 'right' way to do things is to have a branch that you use to track the upstream version, one or more branches that you do development in, and then you create test branches to test merging your work with the upstream work (rather than merging directly into your development or upstream branches). After you test the merge, you delete the branch, you defiantly don't do any more development on it.

only when you are ready to push do you do a merge that will be visible to the outside world and push out.

the excess merge commits that you are bothered by are an indication of real problems, but rebasing instead isn't any better.

pull --rebase

Posted Mar 20, 2012 0:54 UTC (Tue) by dskoll (subscriber, #1630) [Link]

The 'right' way to do things depends on your development team's workflow and the project you're working on. For my main project with only a couple of developers, pull --rebase is almost always the 'right' thing to do.

pull --rebase

Posted Mar 20, 2012 1:10 UTC (Tue) by dlang (✭ supporter ✭, #313) [Link]

quoting Anssi:

> This seems such an obvious change to me, though, that there probably is some good reason why it hasn't been done already..

I was providing the reason why it's not such an obvious change.

pull --rebase

Posted Mar 20, 2012 1:21 UTC (Tue) by Anssi (subscriber, #52242) [Link]

Right. I guess majority of my pushes are smallish single-commit ones, which IMHO do not warrant visible merges (I guess this depends a bit on the project and the workflow, though). For bigger stuff I normally do what you desribed (though often they don't get visible merges either, if I send them via email to upstream who then uses git am).

And after thinking about this a bit more, while I run "git pull --rebase" very often, most times it is on the upstream-tracking branch so "git pull" should actually work just the same.

That is, unless I've made some quick fix commit in the upstream-tracking branch before send-emailing it (when I've forgotten to create a branch for it), in which case "git pull" would do a merge while "git pull --rebase" would usually drop the local one. As said, though, that shouldn't happen :)

pull --rebase

Posted Mar 20, 2012 10:57 UTC (Tue) by adobriyan (guest, #30858) [Link]

> rebasing means that you end up with a new version of code that you have never tested

If rebased code is the same as merged code, yes, it was tested.

pull --rebase

Posted Mar 20, 2012 12:56 UTC (Tue) by nevets (subscriber, #11875) [Link]

> If rebased code is the same as merged code, yes, it was tested.

But that's the big "If". When I first started using git, the upstream maintainer use to rebase my changes into his tree. The problem I had was I never knew what changes he pulled. I would push out a patch series of commits but I never knew what patches actually made it upstream. This changed later when Linus told everyone how to do it, and my push requests were then merged.

The problem is that I can not check what patches I pushed actually made it into upstream. There were some cases that a patch or two was dropped, and I had no clue. It was dropped because of conflicts or someone questioned the patch. But because the other patches were rebased in, I had to go patch by patch to determine what was in and what was not, and this was very time consuming.

I totally understand the usefulness of git pull being a rebase, but it should not be the default. Could you imagine Linux if Linus rebased everyone's pull requests?

pull --rebase

Posted Mar 20, 2012 21:38 UTC (Tue) by mathstuf (subscriber, #69389) [Link]

Unless I'm sorely mistaken, pull --rebase doesn't rebase pulled branches onto another branch. Instead, it rebases any changes that you have locally that are on a branch onto whatever you pulled down from the remote for that branch. The only problem I've encountered with this when I've made merges into master or next and pulling rebases the merge on top of the tree. A pull --rebase should probably pass -p to rebase so that merges are re-merged instead of streamlined on the branch.

pull --rebase

Posted Mar 20, 2012 22:30 UTC (Tue) by nevets (subscriber, #11875) [Link]

You may not be mistaken, I may be. I've never used a pull --rebase before, as that's not part of my workflow.

My workflow would be to do a git remote update of the repo that I'm going to base my work on, and then start adding commits to it. Then push those changes upstream.

The repos I take changes from, I do a pull from others, which I do not want to have any rebase (mine or theirs). If you pull from multiple people, how would this work? One pull would rebase against the other.

If I have old commits and want to rebase them on the latest upstream, I simply do a git remote update of the upstream branch and then do a local rebase.

Now I'm confused to the purpose of pull --rebase?

pull --rebase

Posted Mar 20, 2012 23:14 UTC (Tue) by mathstuf (subscriber, #69389) [Link]

Suppose you have two machines with different commits on some branch. You push from one machine to a common remote repository. On the other machine, you need to pull.

Before:

--B--R (origin/branch)


--B--L (branch)

pull --merge:

--B--R
      \
       \
--B--L--R'

pull --rebase:

--B--R
     | (fast-forward merge on top of B, then rebase L on top of R)
     |
--B--R--L'

The problem is that pull --merge creates a merge commit on the branch with "itself". These commits are, IMNSHO, ugly and shouldn't happen. Now what happens if L is some topic branch that got merged into the branch?

Before:

--B--R (origin/branch)


--B-----L (branch)
       /
      /
--S--T (topic)

pull --rebase (no -p flag):

--B--R
     |
     |
--B--R--S'--T'

pull --rebase (with -p flag):

--B--R
     |
     |
--B--R--L'
       /
      /
--S--T

With the -p flag, pulling from the remote, keeps the repository looking most similar to what it did before the pull which is one reason why I like it more. The other is that I believe that feature branch merges should *always* use --no-ff so that reverting them is easy (reverting L' is easy; S' and T' less so, especially if topic will land later at some point and it's just broken now).

As for multiple remotes sharing a single branch name, all I can say is that I prefer to not do that. I name remotes in namespaces for non-origin repositories (so foobar's github clone is the gh/foobar remote and I checkout foobar's branches as foobar/branchname). Even my forks get named gh/mathstuf on my machines so that if I keep a personal master branch (mathstuf/master) with my feature branches merged in, it doesn't conflict with the actual master branch. My personal feature branches get named under the dev/ namespace and are not under a username. If they become collaborative, then I'd start namespacing them.

pull --rebase

Posted Mar 21, 2012 15:42 UTC (Wed) by misiu_mp (guest, #41936) [Link]

Yes, you should test the code after a rebase as well as after a merge, but, as we all know testing is not always bullet proof. Part of the testing is code review, including the one you do when you modify the existing code.
Take the following history:

A - B - C - D - H
      \       /
        F - G

you implement a feature (F, G) starting from version B.
In order to implement the feature, you have to investigate and understand how the existing code works. When you do that you see version B and the history above accurately depicts it.
Now imagine C introduces a change that affects the code you analyzed in a logical, non-trivial way (that does not introduce a merge conflict with your feature). Your feature will now behave incorrectly, but in a very subtle way. If you based your feature off C or D, you would have probably analyzed the code correctly and made a correct implementation. But you didn't. You maybe just quickly review the changes in C and D, run tests and push to H, merging without conflicts. The tests didn't pick up the subtle error and you've got a bastard-of-a-bug, which depends on a simultaneous modification in two branches, but not in each separately.
If you rebase, the other branch disappears and the history looks as if you indeed started your work after C, indicating you should have seen the changed behavior and its code.

A - B - C - D - F - G

After the bug was found it will be much more difficult for anyone, including you to realize what happened. The lack of rebasing would not prevent the bug, but it would make it easier to track and fix.

pull --rebase

Posted Mar 20, 2012 17:12 UTC (Tue) by njs (guest, #40338) [Link]

> After you test the merge, you delete the branch, you defiantly don't do any more development on it.
> only when you are ready to push do you do a merge that will be visible to the outside world and push out.

I've seen people saying this for a while and it completely baffles me. Maybe you can explain.

What's so bad about merging from time-to-time as you work on your branch? That way you're sticking closer to upstream, can handle any changes and testing incrementally, and don't have to repeat merges over and over. I know it can work -- we did this all the time in monotone (the predecessor to git and hg), and it was wonderful. You end up with extra merges recorded in history, but who cares... they happened... they let future merges work better... instead we have kluges like git-rerere that try to tack a fake merge tool onto a system that already has a real one.

Back in the CVS days the inability to track merges into branches was considered one of the biggest bugs. Now that we fixed it, everyone seems to think that was actually a feature. Why?

pull --rebase

Posted Mar 20, 2012 18:39 UTC (Tue) by dlang (✭ supporter ✭, #313) [Link]

I'm not really a good person to explain the reasons behind this. I'm just passing on what I have learned from watching the discussions on the various discussion lists (especially the kernel and git lists)

I agree it sounds like it should improve development, but as I understand it, the actual result is that it complicates matters because you are pulling in unrelated changes with each merge. These unrelated changes may cause other problems along the way.

I suspect that how big an issue this is depends on the scope of the changes you are making.

If you are working on global-scale changes that touch lots of places, keeping up with upstream changes is more important, and just about all changes are relevant to the work that you are doing.

If you are working on a subsystem that is (relatively speaking) a small portion of the whole, then there are fewer benefits and more drawbacks.

pull --rebase

Posted Mar 21, 2012 16:03 UTC (Wed) by misiu_mp (guest, #41936) [Link]

Here is how I see it: everybody likes a nice linear history. Fat trees with lots of merges are difficult to follow and thus not so useful, which is a good reason to merge less, at least in the public history.
The test-merge way of working is one way to do it: you make modifications, test-merge, make modifications, test-merge, make more modifications and perhaps make a final merge or even rebase.
That would be fine if nothing relevant happened during those merges, so that you could have just as well not made them.
I agree with you though, that relevant merges, where you needed to fix or change something, could (should?) stay in the history.
As usual moderation should be the way to go. The great thing about git is that you have a choice.

pull --rebase

Posted Mar 21, 2012 14:12 UTC (Wed) by slashdot (guest, #22014) [Link]

If you are actively developing the code, rebase is correct because you'll test it against the new upstream code as part of normal development.

If you aren't actively developing it, then you should have it published already, and thus merge is correct.