Dne 26.4.2011 07:35, Robert Haas napsal(a):
> On Apr 13, 2011, at 6:19 PM, Tomas Vondra <tv@fuzzy.cz> wrote:
>> Yes, I've had some lectures on non-linear programming so I'm aware that
>> this won't work if the cost function has multiple extremes (walleys /
>> hills etc.) but I somehow suppose that's not the case of cost estimates.
>
> I think that supposition might turn out to be incorrect, though. Probably
> what will happen on simple queries is that a small change will make no
> difference, and a large enough change will cause a plan change. On
> complex queries it will approach continuous variation but why
> shouldn't there be local minima?
Aaaah, damn! I was not talking about cost estimates - those obviously do
not have this feature, as you've pointed out (thanks!).
I was talking about the 'response time' I mentioned when describing the
autotuning using real workload. The idea is to change the costs a bit
and then measure the average response time - if the overall performance
improved, do another step in the same direction. Etc.
I wonder if there are cases where an increase of random_page_cost would
hurt performance, and another increase would improve it ... And I'm not
talking about individual queries, I'm talking about overall performance.
regards
Tomas