On Fri, Jan 8, 2010 at 2:23 PM, Tom Lane <tgl@xxxxxxxxxxxxx> wrote:
If the other plan does turn out to be faster (and I agree with Tom
that there is no guarantee of that), then one thing to check is
whether seq_page_cost and random_page_cost are set too high. If the
data is all cached, the default values of 4 and 1 are three orders of
magnitude too large, and they should also be set to equal rather than
unequal values.
Tweaking the cost parameters to suit your local situation is the
recommended cure for planner misjudgments; but I'd recommend against
changing them on the basis of only one example. You could easily
find yourself making other cases worse. Get a collection of common
queries for your app and look at the overall effects.
No argument, and well said -- just trying to point out that the
default values really are FAR too high for people with databases that
fit in OS cache.
...Robert
