Re: AW: [HACKERS] Really slow query on 6.4.2

The Hermit Hacker <scrappy(at)hub(dot)org> writes:
> I'm not sure what is all contained in the stats, but the easiest one, I
> think, to have done automagically is table sizes...add a tuple, update the
> table of number of rows automatically. If that numbers gets "off", at
> least it will be more reasonable then not doing anything...no?

The number of tuples is definitely the most important stat; updating it
automatically would make the optimizer work better. The stuff in
pg_statistics is not nearly as important.

The only objection I can think of to auto-updating reltuples is that
it'd mean additional computation (to access and rewrite the pg_class
entry) and additional disk I/O (to write back pg_class) for every INSERT
and DELETE. There's also a potential problem of multiple backends all
trying to write pg_class and being delayed or even deadlocked because of
it. (Perhaps the MVCC code will help here.)

I'm not convinced that accurate stats are worth that cost, but I don't
know how big the cost would be anyway. Anyone have a feel for it?

regards, tom lane