In article <b42b73150804150715r83cad1doa166230ec509f0d@mail.gmail.com>,
"Merlin Moncure" <mmoncure@gmail.com> writes:
> On Mon, Apr 14, 2008 at 5:21 AM, Harald Fuchs <hari.fuchs@googlemail.com> wrote:
>> I think there's something sub-optimal with generate_series.
>> In the following, "documents" is a table with more than 120000 rows,
>> vacuumed and analyzed before the queries.
> everything is working exactly as intended. while it's obvious to you
> that the generate series function returns a particular number of rows
> based on your supplied inputs, it's not (yet) obvious to the planner.
Which was exactly my point. Since generate_series is a builtin
function, the planner could theoretically know the number of rows
returned, thus choosing a better plan.
OTOH, the difference between theory and reality is in theory smaller
than in reality.
> your genser function supplies the hint the planner needs and it
> adjusts the plan. most set returning functions (particularly
> non-immutable ones) are not so easy to determine the # of rows from
> the input parameters anyways.
Yes, of course. I used "genser" just to show that there is a better plan.