Greg Stark <gsstark@mit.edu> writes:
> And a 5% sample is a pretty big. In fact my tests earlier showed the i/o from
> 5% block sampling took just as long as reading all the blocks. Even if we
> figure out what's causing that (IMHO surprising) result and improve matters I
> would only expect it to be 3-4x faster than a full scan.
One way to reduce the I/O pain from extensive sampling would be to turn
VACUUM ANALYZE into a genuine combined operation instead of a mere
notational shorthand for two separate scans.
I'd still be worried about the CPU pain though. ANALYZE can afford to
expend a pretty fair number of cycles per sampled tuple, but with a
whole-table sample that's going to add up.
regards, tom lane