Re: contsel and gist

thanks for the prompt reply.

On Oct 28, 2010, at 10:50 AM, Tom Lane wrote:

>> 1 - am i wrong in my assessment? is the constant contsel, areasel, etc hurting us?
>
> The stub selectivity functions definitely suck.

i'm taking this as implying that my intuition here is basically right.

>> 2 - how hard would it be to implement contsel et al for period data types?
>
> If it were easy, it'd likely have been done already :-(

i am interested in learning more about this, in hopes that it might be possible for me to do it some day. do you have any pointers as far as things to look at to learn from? i imagine this must be a problem for the postgis people too.....

i guess the first step is to figure out what kind of statistics / histograms to collect for the period datatype. (i don't see anything in pg_stats.) has there been previous work / thinking on this?

> However, having said that: the constant value of the stub contsel
> function is intended to be small enough to encourage use of an
> indexscan. Maybe we just need to decrease it a bit more. Have you
> investigated what the cutover point is for your queries?

i'd be happy to investigate this for you, but my guess is my dataset is probably not a good example to use for setting the constant more generally. i'm joining an 8e10 table vs a 150K table, so the selectivity fraction would probably have to drop by many orders of magnitude. that being said, i'll poke around and see if i can find the cutoff point. is there an easy way to do this that doesn't involve recompiling postgres?

best regards, b