postgres_fdw prefers fast plans
Hi.
There's a long-standing issue with postgres_fdw - as it uses cursors, it
prefers plans, optimized for fetching first rows. In bad scenarios this
leads to suboptimal choice of join methods (e.g. choosing nest loop over
hash join) on remote side. I've crafted WIP patch, which tries to fix
this issue. It adds FETCH ALL cursors and uses them in postgres_fdw.
What do you think? Should we go in this direction?
I've looked at attempt to avoid cursors with postgres_fdw [1], but it
seems to avoid dealing with async foreign plans (and looks more
intrusive).
1)
/messages/by-id/CA+FpmFcmO5ctjYgQxSomJC=mCugqPo+51Le2wdxX0kWxjvBBig@mail.gmail.com
--
Best regards,
Alexander Pyhalov,
Postgres Professional
Attachments:
0001-Add-DECLARE-CURSOR-.-FETCH-ALL-and-use-it-in-postgre.patchtext/x-diff; name=0001-Add-DECLARE-CURSOR-.-FETCH-ALL-and-use-it-in-postgre.patchDownload+281-153
On Wed, 2025-12-10 at 15:44 +0300, Alexander Pyhalov wrote:
There's a long-standing issue with postgres_fdw - as it uses cursors, it
prefers plans, optimized for fetching first rows. In bad scenarios this
leads to suboptimal choice of join methods (e.g. choosing nest loop over
hash join) on remote side. I've crafted WIP patch, which tries to fix
this issue. It adds FETCH ALL cursors and uses them in postgres_fdw.
What do you think? Should we go in this direction?
Why not simply set cursor_tuple_fraction to 1.0 in postgres_fdw sessions?
Yours,
Laurenz Albe
Laurenz Albe писал(а) 2025-12-10 16:12:
On Wed, 2025-12-10 at 15:44 +0300, Alexander Pyhalov wrote:
There's a long-standing issue with postgres_fdw - as it uses cursors,
it
prefers plans, optimized for fetching first rows. In bad scenarios
this
leads to suboptimal choice of join methods (e.g. choosing nest loop
over
hash join) on remote side. I've crafted WIP patch, which tries to fix
this issue. It adds FETCH ALL cursors and uses them in postgres_fdw.
What do you think? Should we go in this direction?Why not simply set cursor_tuple_fraction to 1.0 in postgres_fdw
sessions?Yours,
Laurenz Albe
Hi.
Haven't thought about this. Thanks for the idea. The only drawback I see
is that this will affect all cursors, which can be used, for example, by
shippable functions. But looks much simpler.
--
Best regards,
Alexander Pyhalov,
Postgres Professional