On Fri, Jul 1, 2016 at 12:17 PM, Sfiligoi, Igor <Igor.Sfiligoi@ga.com> wrote:
> Hello.
>
> We have a view that is very generic, and we noticed that PostgreSQL is not
> very good at removing useless joins, which makes our queries very slow.
>
> We could change our code to avoid the view and write ad-hoc queries to the
> underlying tables, but would prefer not to, if there is a way around it.
>
> (BTW: We are currently using psql 9.4)
>
> Here is a simplified implementation:
>
> # create table a (id int primary key, name varchar(128));
>
> # create table b (id int primary key, name varchar(128));
>
> # create table c (id int primary key, a_id int references a(id), b1_id int
> references b(id), b2_id int references b(id), b3_id int references b(id));
>
> # create view v as select c.id, c.a_id, c.b1_id, c.b2_id , c.b3_id, a.name
> a_name, b1.name b1_name, b2.name b2_name, b3.name b3_name from c, a, b b1,
> b b2, b b3 where c.a_id=a.id and c.b1_id=b1.id and c.b2_id=b2.id and
> c.b3_id=b3.id;
>
> When I try to get just info from tables c and b1:
>
> # select id, b1_name from v
>
> it still does all the joins (see below).
>
> I would expect just one join (due to the request of columns from the two
> tables),
>
> since all joins are on foreign constrains referencing primary keys,
>
> there are no filters on the other tables,
> so it is guaranteed that the useless joins will always return exactly one
> answer.
I think what you're asking for is a lot more complex than it sounds,
and incorrect. The precise state of the data influences how many
records come back (in this case, either 1 or 0), for example if b3_id
is null you get zero rows. More to the point, you *instructed* the
server to make the join. There are strategies to make joins
'optional' at run time with respect to a query, but they are more
complicated than simply withdrawing columns from the select list.
Stepping back a bit, the query needs to be planned before peeking at
the data in the tables. The planner is able to make assumptions
against a statistical picture of the data but shouldn't be expected to
actually inspect precise result data in order to generate a better
plan.
merlin