Date: Thu, 5 Jul 2018 14:22:32 -0400
From: Alvaro Herrera <alvherre@2ndquadrant.com>
To: Arseny Sher <a.sher@postgrespro.ru>
Cc: Tom Lane <tgl@sss.pgh.pa.us>, pgsql-hackers@lists.postgresql.org
Subject: Re: pgsql: Fix "base" snapshot handling in logical decoding
Message-ID: <20180705182232.3n7ih4kyhtwhmvod@alvherre.pgsql>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <87r2kliqja.fsf@ars-thinkpad>
User-Agent: NeoMutt/20170306-137-4415bd-dirty (1.8.0)
Precedence: bulk

Hi Arseny.  I'm writing a commit message to push this test change, and I
can't explain this bit:

On 2018-Jul-02, Arseny Sher wrote:

> 3) As a side note, answer to my question 'why do we get different errors
>    with VACUUM and VACUUM FULL' is the following. With VACUUM FULL, not
>    only old pg_attribute entry is pruned, but also xmin of new entry
>    with attisdropped=true is reset to frozen xid. This means that
>    decoding session (RelationBuildTupleDesc) actually sees 3 attributes,
>    and the fact that one of them is dropped doesn't embarrass this
>    function (apparently relnatts in pg_class is never decremented) --
>    we just go ahead and decode only live attributes.

I just don't see it that VACUUM FULL would change the xmin of anything
to FrozenXid, and in my experiments it doesn't.  Did you mean VACUUM
FREEZE?

PS - sorry about the broken CC I added :-(

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services