Fujii Masao wrote:
> On Fri, Jun 26, 2009 at 7:14 AM, Heikki
> Linnakangas<heikki.linnakangas@enterprisedb.com> wrote:
>> We certainly update it an order of magnitude more often than before, but
>> I don't think that's an issue. We're talking about archive recovery
>> here. It's not like in normal operation where a corrupt pg_control file
>> means that you lose your data. It will stop the server from starting up,
>> but there's many other files that can be corrupt in a way that causes
>> recovery to fail or stop too early.
>
> Frequent updating of pg_control causes the significant performance
> degradation of archive recovery. I think that this is an issue to be fixed.
> The warm-standby users (including me) care about the performance
> of the standby server, because that affects the failover time, for example.
Are you actually seeing performance degradation caused by frequent
pg_control updates? In the simple test scenarios I've tested, pg_control
is updated only once every few WAL segments, and this with
shared_buffer=32MB. With larger shared_buffers, it happens even less
frequently.
There's a DEBUG2-line in UpdateMinRecoveryPoint() that you can bump to
LOG level if you want to observe that behavior.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com