Re: Push down more full joins in postgres_fdw

On 2017/01/05 21:11, Ashutosh Bapat wrote:
> On Thu, Jan 5, 2017 at 5:14 PM, Etsuro Fujita
> <fujita(dot)etsuro(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>> On 2017/01/03 17:28, Ashutosh Bapat wrote:
>>> In build_subquery_tlists(), why don't we handle base relations?
>>> + if (foreignrel->reloptkind != RELOPT_JOINREL)
>>> + return;

>> The reason for that is we don't need to handle the baserel cases; the tlist
>> for a base relation, if needed, would be created while recursing into a join
>> relation that joins the base relation to other base/join relation.

> Right. Sorry, I misunderstood the code. May be a comment would help.

Will add the comment.

>>> Also, in this function, if fpinfo->tlist is already set, why do we want to
>>> build it again?

>> When this function gets called, fpinfo->tlist isn't set for any base or join
>> relation that needs to build the tlist, so we always need to build it for
>> each such relation.

> IIUC, for a relation with use_remote_estimates we will deparse the
> query twice and will build the targetlist twice.

That's right. We could avoid the duplicate work the way you proposed,
but I was thinking to leave that for another patch. Should we do that
in this patch?

>>> In build_tlist_to_deparse(), if fpinfo->tlist for the given relation is
>>> set, we
>>> should just return it rather than constructing it again.

>> In that function we wouldn't have such cases for base or join relations
>> needing the tlist.

> Same explanation as above.

Will revise if it's better to do that in this patch.

Best regards,
Etsuro Fujita