Thanks everyone. I always thought I *knew* what splice did (replacing elements), and didn't read the perldocs.

Tests on smaller datasets did indeed show that hashes were significantly faster than arrays here. Deleting the hash-element or just setting it to 0 or undef made very little performance difference on 100 MB files. Deleting the element will prevent any problems with duplicates down the road, so I went that way.

Thanks again monks!