Cott Lang <cott@internetstaff.com> writes:
> Fiddling with the above values, only setting sort_mem absurdly large
> easily causes NAN.
Ah. I see an overflow case for sort_mem exceeding 1Gb; that's probably
what you tickled.
I've fixed this in HEAD, but it doesn't seem worth back-patching.
If you care, the change in HEAD is
*** src/backend/optimizer/path/costsize.c.orig Sun Aug 29 01:06:43 2004
--- src/backend/optimizer/path/costsize.c Fri Oct 22 20:02:39 2004
***************
*** 566,572 ****
if (nbytes > work_mem_bytes)
{
double npages = ceil(nbytes / BLCKSZ);
! double nruns = nbytes / (work_mem_bytes * 2);
double log_runs = ceil(LOG6(nruns));
double npageaccesses;
--- 566,572 ----
if (nbytes > work_mem_bytes)
{
double npages = ceil(nbytes / BLCKSZ);
! double nruns = (nbytes / work_mem_bytes) * 0.5;
double log_runs = ceil(LOG6(nruns));
double npageaccesses;
but the variable names have changed since 7.4 so this won't apply
cleanly.
regards, tom lane