From: | Amit Langote <amitlangote09(at)gmail(dot)com> |
---|---|
To: | Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> |
Cc: | Justin Pryzby <pryzby(at)telsasoft(dot)com>, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>, PostgreSQL Hackers <pgsql-hackers(at)lists(dot)postgresql(dot)org> |
Subject: | Re: ALTER TABLE .. DETACH PARTITION CONCURRENTLY |
Date: | 2021-04-26 12:04:55 |
Message-ID: | CA+HiwqFffDsManDTKRaBkg8Y2DkOFDKNe_Wudu_SSX67wuMm3Q@mail.gmail.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
Hi Alvaro,
On Sat, Apr 24, 2021 at 8:31 AM Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
> On 2021-Apr-23, Alvaro Herrera wrote:
> > I think the patch I posted was too simple. I think a real fix requires
> > us to keep track of exactly in what way the partdesc is outdated, so
> > that we can compare to the current situation in deciding to use that
> > partdesc or build a new one. For example, we could keep track of a list
> > of OIDs of detached partitions (and it simplifies things a lot if allow
> > only a single partition in this situation, because it's either zero OIDs
> > or one OID); or we can save the Xmin of the pg_inherits tuple for the
> > detached partition (and we can compare that Xmin to our current active
> > snapshot and decide whether to use that partdesc or not).
> >
> > I'll experiment with this a little more and propose a patch later today.
>
> This (POC-quality) seems to do the trick.
Thanks for the patch.
> (I restored the API of find_inheritance_children, because it was getting
> a little obnoxious. I haven't thought this through but I think we
> should do something like it.)
+1.
> > I don't think it's too much of a problem to state that you need to
> > finalize one detach before you can do the next one. After all, with
> > regular detach, you'd have the tables exclusively locked so you can't do
> > them in parallel anyway. (It also increases the chances that people
> > will finalize detach operations that went unnoticed.)
That sounds reasonable.
> I haven't added a mechanism to verify this; but with asserts on, this
> patch will crash if you have more than one. I think the behavior is not
> necessarily sane with asserts off, since you'll get an arbitrary
> detach-Xmin assigned to the partdesc, depending on catalog scan order.
Maybe this is an ignorant question but is the plan to add an elog() in
this code path or a check (and an ereport()) somewhere in
ATExecDetachPartition() to prevent more than one partition ending up
in detach-pending state?
Please allow me to study the patch a bit more closely and get back tomorrow.
--
Amit Langote
EDB: http://www.enterprisedb.com
From | Date | Subject | |
---|---|---|---|
Next Message | Peter Eisentraut | 2021-04-26 12:10:44 | Re: Dumping/restoring fails on inherited generated column |
Previous Message | vignesh C | 2021-04-26 11:58:55 | Enhanced error message to include hint messages for redundant options error |