Cédric Villemain wrote:
> Le Wednesday 16 April 2008, Magnus Hagander a écrit :
> > Cédric Villemain wrote:
> > > Notice that :
> > >
> > > http://search.postgresql.org/search?q=tom+lane&m=1&l=&d=1&s=r
> > > and
> > > http://search.postgresql.org/search?q=tom+lane&m=1&l=&d=1&s=d
> > >
> > > do not provide same result (3 results by date, 1 by rank) even if
> > > only the sorting is changed.
> >
> > Actually, I get 5 and 7, in the other order.
> >
> > The reason for this is that Tom Lane is way too active. It's
> > gin_fuzzy_search_limit that's doing the restriction first, and the
> > date restriction comes in later.
>
> Yes, you are perfectly right. Can I suggest to deactivate
> gin_fuzzy_search_limit (or increase the value) when one condition
> (and use the condition earlier) can considerably reduce the number of
> results (like the 'post date' here) ?
You'd have to convince the planner to actually not use an indexscan at
all on the tsvector. Normally it'll choose an index scan on each and
then a bitmap join, and we don't want to bring back so many rows...
//Magnus
