My git workflow for anything non-trivial is pretty straightforward, but a little different from what I’ve seen others describe. I’m almost certainly not the only person out there who uses this workflow, so I’m certainly not claiming any originality.
There seem to be two camps, both based on how stuff gets into master, or whatever the equivalent happens to be.
The rebase camp take their feature branches, rebase them off of master, and then perform fast-forward merges, leaving them with a nice, clean, linear history on their master branch.
The merge camp take their feature branches and do a non-fast-forward merge, leaving them with a full history of development.
I think both these camps are wrong. Kind of.
There’s real value to what both camps prize: a linear history makes it much easier to reason about how the master branch got to the state it’s currently in, while merge commits, used well, give you a better idea of when particular features entered the codebase.
My preferred way of working sits in between the two of these. When I’m ready to merge in a branch, I rebase it off of master, run my test suite to make sure everything still works, and then, most importantly, I do a non-ff merge with a proper commit message.
So, I’m essentially both rebasing and merging.
What are the benefits of this?*
Ultimately, you keep your history so that somebody can read it. You hope that this is never necessary, but inevitably the point will come when you have to.
The big benefit of a merge-centric workflow is that you end up with a delimited history: your history is broken up into chunks of changesets that do something in particular to the codebase. This makes things like bisections to find regressions might have been introduced easier. The downside is that you can end up with a spider’s web of a history that’s difficult for somebody to follow.
A rebase-centric workflow doesn’t have this latter problem as a linear history is easy to follow from commit to commit, but you lose the the visibility on when specific things were introduced into the codebase.
This history of when merges happened is actually useful, even if the rat’s nest they cause under a purely merge-centric workflow isn’t quite so useful.
Are there downsides?*
None that I’ve been able to detect, though I have to say that git works against it by not having an easy way to check if a fast-forward merge would work. For this to work, you need to perform a non-fast-forward merge when a fast-forward merge would work. There are ways of doing such a check using git merge-base, git merge-tree, and some other bits and bobs on the output but I need to write up some supporting aliases.