On 08.06.2011 06:37, Kevin Grittner wrote:
> You're right; that one was a false alarm. While REINDEX was reading
> the heap to build the index it got an SIREAD lock on a *heap* page.
We never take locks on heap pages directly, so it must've been a
promotion from heap tuple locks.
> While that could arguably be avoided, even though the heap is being
> read in a serializable transaction, I'm not inclined to get really
> excited about it. If someone wants to dodge it, they can always run
> the REINDEX in READ COMMITTED or REPEATABLE READ mode. Maybe 9.2
> material if there's nothing to do that matters more than that.
That should be pretty easy to avoid, though. I think we should fix it now.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com