It depends on the file system. If possible, you may want to break it down into a couple of directory levels.

• Some file systems choke on large directories (including classical unix directories).
• If you have to scan the directory to find the file that you want (depends, again, on the structure used on the file system), you will end up, on average, with N/2 comparisons to find the file you want. If you break this down into one or two additional levels of subdirectories, it can become a log(N) N^{1/2^d} function instead.

Let's assume that your key column is a 4 digit number (0000-9999) - if you store 100 files in a directory, you would store the files for key 1200-1299 in 1/12/$key.

Update: It is possible that the file system is already a database (IIRC, BeOS had something like this), in which case, you may not have a problem with a large directory. I think that this is typically not the case.

Update 2: See also Re: Efficient processing of large directory, another node that references qmail's use of multi-level directories for its queuing system, and the reasoning behind it.

Update 3: Fixed big-oh of calculations. Thanks JavaFan.