My hypothesis is that the scenarios the SQLite team has setup don’t particularly exploit the fact that the SQLite metrics become more favorable the more files are accessed.
I took a look at the code, and the file based code is basically the worst case for the file system, accessing about as many files as it can.
Additionally, it's doing a lot of unnecessary work. It's using buffered I/O to do a full file read, and throwing away the buffer immediately, which is an expensive way to add an unnecessary memcpy. It's opening by full path instead of openat, which incurs a path lookup penalty.
I think the file based code can be quite a bit faster, if you put a bit of effort in.
I think in serious use, the gap would narrow, not grow.
