MIME-Version: 1.0
References: <CALNXOfqDr5jVRZHThyXyNftDuezjxOJB1mqw=4x1T8YBibi6uw@mail.gmail.com>
 <CAG1ps1w61XD1LxUf3EBCZ=VR3S_orRj_dk6cxmO5KusBVyrgMg@mail.gmail.com>
 <CAAvG1pChk3W9MWkK-_JPyUdThSoJhv4MueH5i8cZS2T_HUP5dA@mail.gmail.com>
 <CAG1ps1xvnTZceKK24OUfMKLPvDP2vjT-d+F2AOCWbw_v3KeEgg@mail.gmail.com>
 <2962669.1724722813@sss.pgh.pa.us> <CAApHDvpRxyw1G2HVkA=PM5Uqh59moOFqPXyMTsEDUadKo2fQTA@mail.gmail.com>
 <2965760.1724724227@sss.pgh.pa.us> <CAApHDvoQRoKHNfTaw+Sn=bZGKVmn8W-5yBPGP7+98mWJz5A9uQ@mail.gmail.com>
 <3104695.1724775341@sss.pgh.pa.us> <3147330.1724795532@sss.pgh.pa.us>
In-Reply-To: <3147330.1724795532@sss.pgh.pa.us>
From: Richard Guo <guofenglinux@gmail.com>
Date: Wed, 28 Aug 2024 11:30:45 +0800
Message-ID: <CAMbWs4_rgwHYAGr=st8XFtdbcKHCf8ti7RpHxJfzv6iZ3AgxJA@mail.gmail.com>
Subject: Re: Significant Execution Time Difference Between PG13.14 and PG16.4
 for Query on information_schema Tables.
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: David Rowley <dgrowleyml@gmail.com>, nikhil raj <nikhilraj474@gmail.com>, 
	pgsql-hackers@lists.postgresql.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: <https://www.postgresql.org/message-id/CAMbWs4_rgwHYAGr%3Dst8XFtdbcKHCf8ti7RpHxJfzv6iZ3AgxJA%40mail.gmail.com>
Precedence: bulk

On Wed, Aug 28, 2024 at 5:52=E2=80=AFAM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I realized that actually we do have the mechanism for making that
> work: we could apply add_nulling_relids to the expression, if it
> meets those same conditions.

I think this should work, as long as we apply add_nulling_relids only
to Vars/PHVs that belong to the subquery in this case, because only
those Vars/PHVs would be nulled by the outer joins contained in the
nullingrels.

> If we go forward with this, I'm thinking about
> changing add_nulling_relids' API contract to say "if target_relid
> is NULL then all level-zero Vars/PHVs are modified", so that we
> don't need that relid set in non-LATERAL cases.

+1.  In LATERAL case, we can always find the subquery's relids in
rcon->relids.  In non-lateral case, any level-zero Vars/PHVs must
belong to the subquery - so if we change add_nulling_relids' API to be
so, we do not need to have rcon->relids set.

Thanks
Richard