-
Tom Lane authored
need be returned. We keep a heap of the current best N tuples and sift-up new tuples into it as we scan the input. For M input tuples this means only about M*log(N) comparisons instead of M*log(M), not to mention a lot less workspace when N is small --- avoiding spill-to-disk for large M is actually the most attractive thing about it. Patch includes planner and executor support for invoking this facility in ORDER BY ... LIMIT queries. Greg Stark, with some editorialization by moi.
d26559db