Le mercredi 21 juillet 2021, 04:25:00 CEST David Rowley a écrit :
> Here the test claims that it wants to ensure that the order by using
> operator(public.<^) is not pushed down into the foreign scan.
> However, unless I'm mistaken, it seems there's a completely wrong
> assumption there that the planner would even attempt that. In current
> master we don't add PathKeys for ORDER BY aggregates, why would that
> sort get pushed down in the first place?
The whole aggregate, including it's order by clause, can be pushed down so
there is nothing related to pathkeys here.
>
> If I adjust that query to something that would have the planner set
> pathkeys for, it does push the ORDER BY to the foreign server without
> any consideration that the sort operator is not shippable to the
> foreign server.
>
> Am I missing something here, or is postgres_fdw.c's
> get_useful_pathkeys_for_relation() just broken?
I think you're right, we need to add a check if the opfamily is shippable.
I'll submit a patch for that including regression tests.
Regards,
--
Ronan Dunklau