Re: [PERFORM] postgres_fdw and column casting shippability - Mailing list pgsql-performance

From Jeff Janes
Subject Re: [PERFORM] postgres_fdw and column casting shippability
Date
Msg-id CAMkU=1yoqZem84+xXzTa=GQnL1k+vaafO2bCBQMukyqh2SQPjA@mail.gmail.com
Whole thread Raw
In response to Re: [PERFORM] postgres_fdw and column casting shippability  (Tom Lane <tgl@sss.pgh.pa.us>)
List pgsql-performance
On Mon, May 15, 2017 at 3:22 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
Jeff Janes <jeff.janes@gmail.com> writes:
> I've tried versions 9.6.3 and 10dev, and neither do what I expected.  It
> doesn't seem to be a planning problem where it thinks the fast plan is
> slower, it just doesn't seem to consider the faster plans as being options
> at all.  Is there some setting to make it realize the cast is shippable?

AFAICS, postgres_fdw doesn't have any knowledge of CoerceViaIO parse
nodes, so it's never going to consider this type of brute-force cast
as shippable.  Normal casts would presumably be shippable if the
underlying function is considered safe.

So then, the secret is to write it like this:

explain analyze select data from remote2 join remote1 on (int8in(textout(remote2.id))  = remote1.id
   where cutoff > 0.9999;

This works to have the join pushed to the foreign side in 9.6, but not before that.

Thanks,

Jeff

pgsql-performance by date:

Previous
From: Tom Lane
Date:
Subject: Re: [PERFORM] postgres_fdw and column casting shippability
Next
From: Adam Brusselback
Date:
Subject: [PERFORM] GIN index not used if created in the same transaction as query