But in general even key-value stores are designed to store sorted/structured data. Not really massive BLOBs with a little meta data sprinkled in to explain it.
You just run into serious performance issues when you try and store/access >2 GB of "junk" data via some clever database management engine, which is really more accustomed to storing tiny little columns of explain-able data you can write clever queries to.
We've wound up in a situation now where we just use databases to store the meta-data and have left the actual data in files, which from a programming point of view makes things harder (since you're hooking into two methods of data access as opposed to one).