Martin Pitt wrote:
Since our Launchpad developers switched from 8.1 to 8.2.3, they often
encounter a situation when the postmaster gets stuck and needs to be
restarted. This happens on various CREATE commands (FUNCTION,
DATABASE, not consistently).
The backtraces show that the process doing the CREATION, another
client connection that InitPostgres(), and the autovacuum daemon all
hang eternally in PGSemaphoreLock(), so I suspect that there is a race
condition of some sort?
So it seems.. :(
I tried to reproduce this with an hour-long run of CREATE/DROP
DATABASE, but without success.
It occurred to me that the process which does the CREATE is in
_bt_vacuum_cycleid(), is that also related to the autovacuum daemon?
I asked the original reporter (Mark Shuttleworth, in CC) to disable
the autovacuum daemon for testing.
_bt_vacuum_cycleid is always called on page splits, but it is related to
vacuums. All vacuums, not to autovacuum in particular. If you're going
to run more tests, I'd suggest running VACUUM in a loop in the
background, to see if that makes the bug occur.
https://launchpad.net/bugs/93042 has symbolic gdb backtraces of all
three processes that are involved.
At this point I am not sure which further information I could provide.
Mark would be happy provide more information and/or directly talk with
you.
Thanks.
I can't see how a BtreeVacuumLock-related deadlock could occur, we never
acquire any other lock while holding it. I'll try to reproduce this on
my laptop..
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com
---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend
