Re: [HACKERS] postgres_fdw bug in 9.6

(2018/01/16 1:47), Robert Haas wrote:
> On Mon, Jan 15, 2018 at 3:09 AM, Etsuro Fujita
> <fujita.etsuro@lab.ntt.co.jp>  wrote:
>> Yeah, but I don't think the above example is good enough to explain that,
>> because I think the bar/baz join would produce at most one tuple in an EPQ
>> recheck since we would have only one EPQ tuple for both bar and baz in that
>> recheck, and the join is inner.  I think such an example would probably be
>> given e.g., by a modified version of the SQL where we have a full join of
>> bar and baz, not an inner join.
>
> Hmm, I was thinking that bar and baz wouldn't be constrained to return
> just one tuple in that case, but I'm wrong: there would just be one
> tuple per relation in that case.  However, that would also be true for
> a full join, wouldn't it?

Consider:

postgres=# create table bar (a int, b text);
postgres=# create table baz (a int, b text);
postgres=# insert into bar values (1, 'bar');
postgres=# insert into baz values (2, 'baz');
postgres=# select * from bar full join baz on bar.a = baz.a;
  a |  b  | a |  b
---+-----+---+-----
  1 | bar |   |
    |     | 2 | baz
(2 rows)

Both relations have one tuple, but the full join produces two join 
tuples.  I think it would be possible that something like this happens 
when executing a local join plan for a foreign join that performs a full 
join remotely.

> Regardless of that, the patch fixes the reported problem with very
> little code change, and somebody can always improve it further later.

Agreed.

Best regards,
Etsuro Fujita