I recently did experiments with storing documents (blobs) with UUIDs as files. I stored documents within a hierarchy of a few levels of directories to avoid having too many files in a single directory (UUID abcdefg was stored as a/b/c/d/abcdefg). Performance was surprisingly low. Reading documents resulted in a lot of disk seeks. Also it seems that directory entries lookup is a O(n) operation in a standard ext4. I also did experiments with simple databases like GDBM. The effect was also disappointing. In the end I started to really appreciate RDBMSes design and all theory that is behind them. Creating a highly performant system using I/O operations is HARD.