Neil Conway <nconway(at)klamath(dot)dyndns(dot)org> writes:
> On Thu, 2002-02-21 at 22:15, Tom Lane wrote:
>> I believe that SQL-language functions don't have any such optimization
>> at all :-( ... I think they are re-parsed and re-planned from scratch
>> on each call.
> Would it be possible to enhance SQL-language functions to cache their
> query plan?
Certainly; if plpgsql can do it, so can SQL functions. You could even
steal (or better, find a way to share) a lot of the code from plpgsql.
But no one's gotten around to it.
A related improvement that's been in the back of my mind for awhile
is to "inline" trivial SQL functions. If you look in pg_proc you'll
find quite a few SQL functions that are just "SELECT
some-arithmetic-expression". I would like to get the planner to expand
those into the parse tree of the invoking query, so that the function
call overhead goes away completely. For example, bit_length(text) is
defined as "select octet_length($1) * 8", so
SELECT bit_length(f1) FROM text_tbl WHERE ...
could be expanded to
SELECT octet_length(f1) * 8 FROM text_tbl WHERE ...
which seems to run about three or four times as fast (though of course
some of that differential would go away given caching of SQL-function
I don't believe this would be excessively difficult, but it hasn't
gotten to the top of the to-do queue...
regards, tom lane
In response to
pgsql-hackers by date
|Next:||From: Bruce Momjian||Date: 2002-02-22 04:13:21|
|Subject: Re: Summary of new configuration file and data directory locations|
|Previous:||From: Bruce Momjian||Date: 2002-02-22 04:07:10|
|Subject: Re: elog() proposals|