Re: BUG #7572: virtualxid lock held by bgwriter on promoted slaves

On 27 September 2012 22:29, Simon Riggs <simon(at)2ndquadrant(dot)com> wrote:
> On 26 September 2012 22:33, <daniele(dot)varrazzo(at)gmail(dot)com> wrote:
>> The following bug has been logged on the website:
>>
>> Bug reference: 7572
>> Logged by: Daniele Varrazzo
>> Email address: daniele(dot)varrazzo(at)gmail(dot)com
>> PostgreSQL version: 9.1.4
>> Operating system: Linux
>> Description:
>>
>> Hello,
>>
>> when a slave is promoted, the pgwriter keeps holding a lock with virtualxid
>> "1/1" and virtualtransaction "-1/0". Such lock stops pg_reorg to run (as
>> reported here:
>> http://pgfoundry.org/tracker/index.php?func=detail&aid=1011203&group_id=1000411&atid=1376)
>> but I've verified the same condition on 9.1.4).
>>
>> Is it possible to free that lock on slave promotion?
>>
>> Is it safe to ignore that lock for pg_reorg sake? The program is which is
>> probably waiting for all the transactions to finish before swapping the
>> table in the system catalog but I'm not sure about that yet.
>
> This one is mine.... I don't think its important we hold that lock,
> but will check.

This was an interesting bug to track down and certainly had me stumped
for some time.

Nothing at all todo with the WALwriter. It turns out that the Startup
process never ended its VirtualTransaction, which caused 2 separate
bugs in 9.0/9.1 and 9.2/HEAD. The bugs in 9.2+ were only visible
because of lack of initialisation of the fp fields, also fixed.

Thanks for the bug report.

--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services