Re: postgres_fdw join pushdown (was Re: Custom/Foreign-Join-APIs)

On 2016/02/16 16:40, Etsuro Fujita wrote:

On 2016/02/16 16:02, Ashutosh Bapat wrote:

On Tue, Feb 16, 2016 at 12:26 PM, Etsuro Fujita
<fujita.etsuro@lab.ntt.co.jp <mailto:fujita.etsuro@lab.ntt.co.jp>> wrote:

    On 2016/02/16 15:22, Ashutosh Bapat wrote:

        During join planning, the planner tries multiple combinations of
        joining
        relations, thus the same base or join relation can be part of
        multiple
        of combination. Hence remote_conds or joinclauses will get linked
        multiple times as they are bidirectional lists, thus breaking
        linkages
        of previous join combinations tried. E.g. while planning A join
        B join C
        join D planner will come up with combinations like A(B(CD)) or
        (AB)(CD)
        or ((AB)C)D etc. and remote_conds from A will first be linked
into
        A(B(CD)), then AB breaking the first linkages.

    Exactly, but I don't think that that needs to be considered because
    we have this at the beginning of postgresGetGForeignJoinPaths:

         /*
          * Skip if this join combination has been considered already.
          */
         if (joinrel->fdw_private)
             return;

There will be different joinrels for A(B(CD)) and (AB) where A's
remote_conds need to be pulled up.

Agreed.

The check you have mentioned above
only protects us from adding paths multiple times to (AB) when we
encounter it for (AB)(CD) and ((AB)C)D.

Sorry, I don't understand this fully.

Another thing I don't really understand is why list_copy is needed in the second list_concat for the case of INNER/FULL JOIN or in both list_concats for the case of LEFT/RIGHT JOIN, in your patch.  Since list_concat is nondestructive of its second argument, I don't think list_copy is needed in any such list_concat.  Maybe I'm missing something, though.