I try to make each commit a snapshot of a working repository (compileable or runnable or testable, whatever is the heuristic for working) where the difference between each snapshot can be explained by the commit message. Ideally they are isolated to a single logical "unit" of change (fix, refactor, add, remove). The goal here being to minimize the amount of confusion and work for anyone traveling up and down the tree. I often have to rewrite my local history to make this happen, because the actual changes that I make can happen in a somewhat arbitrary order. How has local history revision bitten you?

Rebasing leads to either having long stretches of non-compiling commits in history or giant non-bisectable commits. E.g. you added a method call on your feature branch, that method was renamed in master while you were working on your feature. If you merge then your commits still compile and I can use automated `git bisect` the way it's intended. If you rebase then your commits don't compile and I can't bisect through commits on your feature branch. If you squash then your whole feature development becomes a single monster commit and I can't bisect through it.

I agree with having as many commits as possible be compilable, but that's not the sole criterion, because there's a tension between that and having granular history: if you squash the whole history of the repo into a single commit then that means 100% of commits are compilable, but it's still a bad move. Conversely, a non-compiling commit in between two compiling commits is not a big problem (you just make sure your git bisect script skips non-compiling commits) - what really matters is keeping the diff between two successive compiling commits as small as possible. IME the best way to achieve that is never rewriting history.

That's a very good point with rebasing. Thanks for explaining.