I looked into the behavior complained of here:
http://archives.postgresql.org/pgsql-performance/2012-04/msg00093.php
The problem query can be abstracted to
select * from a, b, c, dwhere a.x = b.y and (a.z = c.c or a.z = d.d)
Table a is much larger than the others (in fact, in the given example
c and d are known to be single rows), and there are indexes on the
mentioned columns of a. In this situation, the best plan is to
cross-join c and d, then use a BitmapOr indexscan to pick out the rows
of a that satisfy the OR condition, and finally join that small number
of rows to b. The planner will use a cross-join-first plan if we omit
b and the first WHERE clause from the query; but in the query as given,
it fails to discover that plan and falls back on a vastly inferior plan
that involves forming the a/b join first.
The reason for this behavior is the anti-clauseless-join heuristics in
join_search_one_level(). Without b, there are no join clauses available
at join level 2, so the planner is forced to form all three 2-way cross
joins; and then at level 3 it finds out that joining a to c/d works
well. With b, we find the a/b join has a usable join clause so we form
that join, and then we decide not to make any 2-way clauseless joins.
So the c/d join is never constructed and there is no way to exploit the
desirable indexscan at higher levels.
After some reflection I think that the blame should be pinned on
have_relevant_joinclause(), which is essentially defined as "is there
any join clause that can be evaluated at the join of these two
relations?". I think it would work better to define it as "is there any
join clause that both these relations participate in?". In the majority
of real-world queries, join clauses relate exactly two relations, so
that these two definitions are equivalent. However, when we do have
join clauses involving 3 or more relations, such as the OR clause in
this example, it's evidently useful to consider cross-product joins of
the smaller relations so that the join clause can be applied during the
scan of the largest table.
It would probably not be a good idea to back-patch such a change, since
it might have consequences I can't foresee at the moment. But I'm
strongly tempted to squeeze it into 9.2. Thoughts?
regards, tom lane
