Please, please, please, don't base it on Windows. Drive letters are a seriously braindead way to map partitions.

>>2
BeOS had a database backend, so technically that's been seen before. :)

Personally I'd like to see a copy-on-write filesystem that consolidates identical file data into symlinks. (unless you specify otherwise; for various mostly performance-related reasons, it's useful to have two identical files, but generally it's a mistake) Also would be neat is the ability to separate the ideas of directories and filenames from the file content -- have a tag-based system underneath this, so instead of say, putting a file in /usr/bin, it would be tagged as 'usr', 'bin', and '/usr/bin'. That would essentially remove the necessity to maintain a $PATH: you'd just grab all directories tagged with 'bin'. This also applies for linked library resolution, help searching, etc. and would generally make the entire system heaps more flexible.

Oh and all that stuff is way easier ... if you put the whole filesystem into a database! BeOS didn't quite go that far, but I imagine if it'd been designed after the whole Web 2.0 thing, it would have used a similar structure. When it was made the whole tagging concept hadn't crossed most people's minds yet and I don't think anyone realized the magnitude of what such a system could allow.