Horton

Re: Concurrent ALTER SEQUENCE RESTART Regression

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: Peter Eisentraut <peter.eisentraut@2ndquadrant.com>
Cc: Michael Paquier <michael.paquier@gmail.com>, Jason Petersen <jason@citusdata.com>, PostgreSQL mailing lists <pgsql-bugs@postgresql.org>
Date: 2017-05-02T15:41:48Z
Lists: pgsql-bugs, pgsql-hackers 
On 2017-05-02 11:05:38 -0400, Peter Eisentraut wrote:
> On 4/27/17 01:52, Andres Freund wrote:
> > In contrast to <v10, the actual change of the tuple is *not* happening
> > with the page lock held.  But now we do log XLOG_SEQ_LOG, then unlock
> > the buffer, and then do a CatalogTupleUpdate().  How is that correct?
> 
> The change to the sequence data and the change to the catalog are two
> separate operations.  There is no need AFAICT for the latter to be done
> while the former is locked or vice versa.

You snipped the salient part of my response:

> Imagine two of these running concurrently - you might end up with
> A:XLogInsert B:XLogInsert B:CatalogTupleUpdate A:CatalogTupleUpdate

Which'll lead, yet another avenue, to sequence states that aren't in
sync with the catalog.

- Andres

Commits

  1. Make ALTER SEQUENCE, including RESTART, fully transactional.

  2. Modify sequence catalog tuple before invoking post alter hook.

  3. Use weaker locks when updating pg_subscription_rel

  4. Add pg_sequence system catalog

  5. Modify sequence state storage to eliminate dangling-pointer problem