Bruce Momjian <maillist@candle.pha.pa.us> writes:
> The thing I liked about the existing algorithm is that it did the
> sorting without the page faulting/thrashing caused by many sort
> algorithms.
Yes, I think we want to stick with the basic architecture of an
initial replacement-selection filter operating in limited memory,
followed by merge passes. I'm just thinking about how to reduce
the amount of disk space chewed up by intermediate merge results...
regards, tom lane