Re: Problem with ALTER TABLE - occasional "tuple concurrently updated"

On Thu, Nov 18, 2010 at 1:35 PM, Alvaro Herrera
<alvherre@commandprompt.com> wrote:
> Excerpts from Alvaro Herrera's message of jue nov 18 15:31:16 -0300 2010:
>> Excerpts from Robert Haas's message of jue nov 18 15:11:37 -0300 2010:
>>
>> > In the current master branch, it appears that "ALTER TABLE c INHERIT
>> > p" takes a ShareUpdateExclusiveLock on the child, which seems
>> > sufficient, and an AccessShareLock on the parent, which seems like it
>> > might not be; though I'm having a hard time figuring out exactly when
>> > it wouldn't be, especially since in 8.4 I'm fairly sure any ALTER
>> > TABLE command takes an AccessExclusiveLock.
>>
>> What if two of these run at the same time, and the parent doesn't
>> have children when they start? =A0They would both try to set
>> relhassubclass, no?
>
> Yep, duplicated the issue that way.

I think ATExecAddInherit() and MergeAttributes() need to take
ShareUpdateExclusiveLock instead of AccessShareLock to prevent this.

--=20
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company