What about when they accidently release private information because the information was at some point in the software. I'd rather a single backup system like time machine than every application reinventing the wheel with all the bugs that entails.
That's why it should be the file system storing backups, not the document itself.
One should also consider a logarithmic decrease in backups: while I'm writing, a backup of the state a few seconds ago is useful, but I don't need every second from weeks ago. Say the system keeps a backup of every few seconds for the last few minutes, every few minutes for the last few hours, every few hours for the last few days, every few days for the last few weeks, every few weeks for the last few months, every few months for the last few years--the odds are then pretty good that I'd be able to revert to a version I find useful, but not requiring an awful lot of versions.
Back of the envelope: 5 versions/minute for 5 minutes = 25 versions; 4 versions/hour for 8 hours = 24 versions; 3 versions/day for seven days = 21 versions; 4 versions/week for four weeks = 16 versions; 1 version/month for 4 years = 48 versions, so a grand total of 134 previous versions of a document, which really isn't that much space, particularly assuming efficient differencing algorithms.
