Collator

The IBM 077 appeared in 1937 with two input hoppers and four output pockets. It read one card from each deck, compared the columns the operator had selected, and dropped the lower card into the pocket before reading the next — precisely the inner loop of a merge, a couple of decades before anyone wrote it down as an algorithm. Two sorted decks in, one sorted deck out, a couple of hundred cards a minute.

The second use was matching. With a master deck in one hopper and the day's transactions in the other, the machine separated hits, misses and duplicates into their own pockets, which amounts to updating a database that does not exist. Paired with the sorter, the collator gave you external sorting of datasets far larger than any machine could hold at once — the same division of labour a modern database index relies on, except that the intermediate results were carried between machines by hand.