On Sat, Nov 7, 2009 at 9:32 PM, Melvin Davidson <melvin6925 at yahoo.com>wrote:

> >Yep, pg 8.3.7 and slony 1.2.14 at the moment.  If I do this on a
> >machine with JUST the objects I'm replicating, I can get 10 seconds
> >per table and 5 or so per sequence.
>
> >On the DB with 29000 other objects not replicated, it gets up to 30
> >seconds per table and 20 seconds per sequence.  Which means a 5 hour
> >time for create set.  And if autovac kicks in it blocks create set.
> >so I had to turn that off.
>
> This sound more like O/S and hardware tuning than anything else.
> Just out of curiosity, what is the O/S and hardware configuration?
>
> I would at least hope that the data and WAL files are on separate disks?
>

This an 8 core opteron 2.1GHz machine with 32G ram and a 12 disk RAID-10 for
pgdata and a two disk RAID-1 for pg_xlog.  The autovac when it kicks off,
since it's set to have a 20 ms cost delay, purposely takes a while, so as
NOT to get in the way.  So, on a 2 to 5Gig table file it can take 10 to 20
minutes.  While this is happening, setaddtable is blocked.  While it's
blocked, all access to the table by other queries are blocked, and nothing
gets though.  This is NOT a tuning issue on the OS / hardware.  This machine
can handle 10 to 20k transactions per minute under normal load.  But the
interaction between autovac and create set / setaddtable halts all access to
the table being vacuumed.  The pg_lock and pg_stat_activity show it pretty
clearly.

However, I can turn off autovac during the create set.  But the create set
takes wayyyy too long.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.slony.info/pipermail/slony1-general/attachments/20091107/=
d48c37f9/attachment.htm