merge vs rebase: Is visualization in gitk the only problem?

There are some discussions going on about merge vs. rebase. The suggestions by Dscho is to prefer rebase over merge in a CVS-style workflow.

Rebase has definitely benefits but not all of its details are obvious at a first glance. Tell a newbie to read the git rebase man page and explain what git reabase does and you know what I mean. Rebase definitely can help to create a cleaner history. But it rewrites history and therefore destroys information, for example information about the original code base a patch was developed against, or merge conflicts that were resolved. You also need to decide when to use rebase and when to use merge. So you need to make a choice.

Why not always use git merge?

Is the only problem of git merge that it might create loops in the history with potentially long running parallel lines that are insufficiently visualized in gitk?

If so, why not improve the visualization?

Or is there any other deficiency of always using merge and never using rebase that I don't see?

Obviously you can use git merge only if you want to have _all_ changes from the other branch. But this is often what you want. In a CVS-style workflow you want to merge all changes from the shared branch before pushing. Why going through the hassel of git rebase?

I'm aware that you can do other things with git rebase that cannot be achieved by using git merge. For example moving a topic branch to a different stable branch without merging all of the topic branch's history ("--onto" option). It is easy to explain that rebase has it's own value.

	Steffen

On 10/27/07, Steffen Prohaska <prohaska@zib.de> wrote:

I'd use git-rebase for when I'm sending stuff upstream that I don't want to force the maintainer to merge, because I can probably do it better and quicker by rebasing. Once that's done, the upstream maintainer can just do a git-am (or similar), and it'll apply neatly on top of the current head. In other words, git-rebase allows the "merge effort" to be shifted to the brancher and away from the mainline maintainer/developer.

Dave.

On Oct 27, 2007, at 10:16 AM, David Symonds wrote:

That definitely makes sense.

But you may need to do more merging because you need to merge every single commit. It may be easier to merge only the final state of a topic branch.

Could we somehow send rerere information together with a patch? This would give the upstream maintainer (and everyone else who wants to test the patch) the help needed to do a merge.

	Steffen

Hi,

On Sat, 27 Oct 2007, Steffen Prohaska wrote:

IMHO this is not sensible to send over email. But think I that something like this could be done for pullers: they fetch from somewhere, and then recreate the rerere information from the merge commits they just got (this would have to be a new option to git-rerere, methinks).

Ciao, Dscho

On Sat, 27 Oct 2007, Steffen Prohaska wrote:

Not all possible upstream systems support non-linear history. If upstream is actually svn or cvs, you need to rebase in order for the other system to be able to represent the result.

I think it's easier for someone from the future trying to figure out the historical reason for some code to be a certain way if that code isn't touched by a merge conflict resolution. If there is a conflict, the resulting content is most directly due to a commit that has no explanation aside from a note that there was a conflict there; it's somewhat easier to understand if the code is entirely attributable to the commit that logically introduces the change. (Although maybe we could make the display take care of this; linearize the history using the rerere info in the merge commit, and show that as the history, annotated suitably to indicate that the blame for some of it falls partially on the merge commit.)

It's also more familiar for many users to be asked to place their changes after all other committed changes, rather than leaving the ancestry pegged at the last time they updated before starting to work.

Unless you've got a series of changes to a section that the upstream also changed, it's the same amount of resolution, but rebase lets you resolve logical collections of conflicts (i.e., those due to a single one of your commits) individually. IIRC, merging my builtin-fetch topic after 1.5.3 had 3 scattered conflicts, and rebasing had 4 conflicts, with no more than one per commit (so that I could test the resolutions independantly), and the rebase was easier.

	-Daniel
*This .sig left intentionally blank*