On Tue, 2009-02-03 at 13:50 +0000, Gregory Stark wrote:
> Hannu Krosing <hannu@krosing.net> writes:
>
> > Actually we came up with a solution to this - use filesystem level
> > snapshots (like LVM2+XFS or ZFS), and redirect backends with
> > long-running queries to use fs snapshot mounted to a different
> > mountpoint.
>
> Uhm, how do you determine which snapshot to direct the backend to? There could
> have been several generations of tuples in that tid since your query started.
> Do you take a snapshot every time there's a vacuum-snapshot conflict and
> record which snapshot goes with that snapshot?
The most sensible thing to do seems to wait for some configurable period
(say a few seconds or a few minutes), delaying WAL apply, and then to do
the snaphot, mount it and redirect all running transactions to use
_current_ filesystem snapshot, and then resume WAL application.
As each of the transactions running on saved fs snapshots complete, they
are switced back to main/live fs view.
When the last there transaction ends, the snapshot is unmounted and
released.
--
------------------------------------------
Hannu Krosing http://www.2ndQuadrant.com
PostgreSQL Scalability and Availability Services, Consulting and Training