been there, done that, shit seriously sucked
you should never use a vcs that has no ability to ask "is this repo in a valid state", or you may try to checkout the point version released to a large customer and be unable to do so. Data corruption can happen, even on raid drives (we're pretty sure it was a bug in the controller, but that doesn't change the effect.)
(I'd also say that atomic commits are an excellent reason to leave CVS behind, but that can be argued.)