Re: lifetime of the old CTID

Christophe Pettus <xof@thebuild.com>

From: Christophe Pettus <xof@thebuild.com>
To: Matthias Apitz <guru@unixarea.de>
Cc: Laurenz Albe <laurenz.albe@cybertec.at>, pgsql-general@lists.postgresql.org
Date: 2022-07-06T05:44:23Z
Lists: pgsql-general

> On Jul 5, 2022, at 22:35, Matthias Apitz <guru@unixarea.de> wrote:
> Internally, in the DB layer, the read_where() builds the row list matching
> the WHERE clause as a SCROLLED CURSOR of
> 
>    SELECT ctid, * FROM d01buch WHERE ...
> 
> and each fetch() delivers the next row from this cursor. The functions
> start_transaction() and end_transaction() do what their names suggest and
> rewrite_actual_row() does a new SELECT based on the ctid of the actual row
> 
>    SELECT * FROM d01buch WHERE ctid = ... FOR UPDATE
>    ...
>    UPDATE ...

On first glance, it appears that you are using the ctid as a primary key for a row, and that's highly not-recommended.  The ctid is never intended to be stable in the database, as you have discovered.  There are really no particular guarantees about ctid values being retained.

I'd suggest having a proper primary key column on the table, and using that instead.