Gary Doades <gpd@gpdnet.co.uk> writes:
> Interestingly, if I don't delete the table after a run, but just drop
> and re-create the index repeatedly it stays a pretty consistent time,
> either repeatedly good or repeatedly bad!
This is consistent with the theory of a data-dependent performance
problem in qsort. If you don't generate a fresh set of random test
data, then you get repeatable runtimes. With a new set of test data,
you might or might not hit the not-so-sweet-spot that we seem to have
detected.
regards, tom lane