Re: Performance problem in PLPgSQL

Marc Cousin <cousinmarc@gmail.com> writes:
> On 23/08/2013 23:55, Tom Lane wrote:
>> My previous suggestion was to estimate planning cost as
>> 10 * (length(plan->rangetable) + 1)
>> but on reflection it ought to be scaled by one of the cpu cost constants,
>> so perhaps
>> 1000 * cpu_operator_cost * (length(plan->rangetable) + 1)
>> which'd mean a custom plan has to be estimated to save a minimum of
>> about 5 cost units (more if more than 1 table is used) before it'll
>> be chosen.  I'm tempted to make the multiplier be 10000 not 1000,
>> but it seems better to be conservative about changing the behavior
>> until we see how well this works in practice.
>> 
>> Objections, better ideas?

> No better idea as far as I'm concerned, of course :)

> But it is a bit tricky to understand what is going on when you get
> hit by it, and using a very approximated cost of the planning time
> seems the most logical to me. So I'm all for this solution.

I've pushed a patch along this line.  I verified it fixes your original
example, but maybe you could try it on your real application?
http://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=005f583ba4e6d4d19b62959ef8e70a3da4d188a5
        regards, tom lane