Re: Group-count estimation statistics

Kris Jurka <books(at)ejurka(dot)com> writes:
> The proposed change biases towards a hash plan which has no provision for
> spilling to disk. Slow is one thing, but excessive memory usage and
> possibly failing is another thing.

Keep in mind that we are replacing 7.4 code that had a serious tendency
to select hash plans when it really shouldn't, because of underestimated
table sizes. Now that we have the physical-size-driven estimate of
table rowcounts, I think we've gone too far over in the other direction.

Which is not to say that spill-to-disk logic wouldn't be a nice thing to
add, but I'm not going to panic about its not being there today. 7.4
presented a much greater hazard than we have now, but we got through
that cycle without a large number of complaints. There's always the
enable_hashagg switch if you really find yourself backed into a corner.

regards, tom lane