From: Tom Lane <tgl@sss.pgh.pa.us>
To: Amit Langote <amitlangote09@gmail.com>
cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Subject: Re: unsupportable composite type partition keys
In-reply-to: 
 <CA+HiwqGug69_zhRCzENJPg_Ynpx_Eb-2-=N1u2Zw3X6T1S3wbQ@mail.gmail.com>
References: 
 <CA+HiwqFUzjfj9HEsJtYWcr1SgQ_=iCAvQ=O2Sx6aQxoDu4OiHw@mail.gmail.com>
 <5361.1576602769@sss.pgh.pa.us>
 <CA+HiwqGiGBgFFcL6YGKRqfX1-gYKmuLLG_JwU_UaSaX2J+jehQ@mail.gmail.com>
 <28160.1576676326@sss.pgh.pa.us>
 <CA+HiwqENLY2ta0goA0jcfk4cWdj2W0r6Ywjf=PQiMpJkyXygmg@mail.gmail.com>
 <26766.1576947735@sss.pgh.pa.us> <24653.1576962784@sss.pgh.pa.us>
 <CA+HiwqGug69_zhRCzENJPg_Ynpx_Eb-2-=N1u2Zw3X6T1S3wbQ@mail.gmail.com>
Comments: In-reply-to Amit Langote <amitlangote09@gmail.com>
	message dated "Mon, 23 Dec 2019 18:42:36 +0900"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <25382.1577112556.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Mon, 23 Dec 2019 09:49:16 -0500
Message-ID: <25383.1577112556@sss.pgh.pa.us>
Precedence: bulk

Amit Langote <amitlangote09@gmail.com> writes:
> On Sun, Dec 22, 2019 at 6:13 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> +             * To ensure that it's not leaked completely, re-attach it =
to the
> +             * new reldesc, or make it a child of the new reldesc's rd_=
pdcxt
> +             * in the unlikely event that there is one already.  (See h=
ack in
> +             * RelationBuildPartitionDesc.)

> While I can imagine that RelationBuildPartitionDesc() might encounter
> an old partition descriptor making the re-parenting hack necessary
> there, I don't see why it's needed here, because a freshly built
> relation descriptor would not contain the partition descriptor after
> this patch.

Well, as the comment says, that's probably unreachable today.  But
I could see it happening in the future, particularly if we ever allow
partitioned system catalogs.  There are a lot of paths through this
code that are not obvious to the naked eye, and some of them can cause
relcache entries to get populated behind-your-back.  Most of relcache.c
is careful about this; I do not see an excuse for the partition-data
code to be less so, even if we think it can't happen today.

(I notice that RelationBuildPartitionKey is making a similar assumption
that the partkey couldn't magically appear while it's working, and I
don't like it much there either.)

>> * It'd be better to declare RelationGetPartitionKey and
>> RelationGetPartitionDesc in relcache.h and get their callers out of the
>> business of including rel.h, where possible.

> Although I agree to declare them in relcache.h, that doesn't reduce
> the need to include rel.h in their callers much, because anyplace that
> uses RelationGetPartitionDesc() is also very likely to use
> RelationGetRelid() which is in rel.h.

Hm.  Well, we can try anyway.

> [1] https://www.postgresql.org/message-id/CA%2BHiwqFucUh7hYkfZ6x1MVcs_R2=
4eUfNVuRwdE_FwuwK8XpSZg%40mail.gmail.com

Oh, interesting --- I hadn't been paying much attention to that thread.
I'll compare your PoC there to what I did.

			regards, tom lane