State of CVS-to-git conversion tools (Was: Re: cvsps: bad usage: invalid argument --norc)

Ilya Basin <basinilya@gmail.com>:

Or you could just use 3.x directly rather than through the git-cvsimport wrapper. It works better that way - it actually ships a fast-import stream that git fast-import can consume directly.

Old cvsps (2.x) was very, very broken; there was a bug in it that pretty much guaranteed incorrect conversion of branchy repos. I've fixed that particular bug, and several other serious ones, along with adding the import-stream output stage, but I don't really trust the cvsps code; I think its algorithmic core is weak and klugey and only works semi-by-accident.

I do *not* recommend downgrading, it will pitch you from a bad situation into a worse one. Yes, Junio is still shipping a wrapper for 2.x, but that was very much against my advice.

I'm also now the maintainer of cvs-fast-export, which used to be Keith Packard's utility for lifting the X.org CVS repo to git. That was before I resurrected it and added a fast-import stream output stage.

I *think* cvs-fast-export's algorithms are more correct than cvsps's, but I have not yet been able to pry loose the time to write the really rigorous test suite to verify this. That goal has been second on my prority list for a couple of months now; I keep wanting to get to it but having to fight fires on other projects instead.

I wish I could point you at a convertor I really trust. I can't. There is a third tool, cvs2git (based on the analyzer from cvs2svn) that I don't maintain, which has problems of its own. And those three are about it.

Yes, it's a swamp. The relatively poor capability of the tools isn't anybody's fault; the problem domain is *nasty*. I've been working a closely related but easier one (Subversion stream dump analysis) for a couple years now and understanding it doesn't make it less ugly.

I think with 4-6 weeks of concentrated attention I could clean up the mess and deliver a really high-quality converter, and I'm motivated to do this because I want it as a CVS front end for reposurgeon. But it hasn't happened yet and the incompleteness of my test suite is a blocker in the way.

The topo analysis code in all these tools is really fragile and tends to break on old edge cases when you try to teach it to handle new ones, so a good set of regression tests is especially important. And doesn't yet exist, though I have built a decent start for cvsps based on the tests in the git tree.

Do you have enough interest and spare cycles to help finish the test suite? Another pair of hands on it might speed things up a lot.