Reading a file on a local filesystem is faster than reading from a database. Maybe they did it that way so they could share code between machines, but that's better done with NFS or rsync in my opinion. The compile step sounds annoying, as does having to restart when changes happen.

When templates live on the filesystem, changes can be automatically picked up by the running system just by stat'ing the file. That could be turned into a database check in their custom Provider class, but I suspect they did this to keep shared memory high. In my experience though, compiling the templates before forking didn't really decrease the amount of unshared memory.