Markus Schaber <schabi@logix-tt.com> writes:
> Tom Lane wrote:
>> The trick is to figure out what a useful parameterized cost model would
>> look like. IIRC, the main reason the xfunc code rotted on the vine was
>> that its cost parameters didn't seem to be either easy to select or
>> powerful in predicting actual cost. We'd have to do better this time.
> I don't know what the xfunc people did, but at least for some varlen
> data types (Arrays, PostGIS, text), some function costs (concatenation,
> GeomUnion etc.) can be estimated via the average field size of the tables
> Has that idea been considered?
[ shrug... ] Concatenation is definitely not one of the functions we
need to worry about. In fact, I'd say that only functions containing
database accesses are really worth worrying about --- and the hard
part of modeling them is the possible variation in the number of rows
touched depending on their parameters.
regards, tom lane