[KLUG Members] Postgres Cluster

[Jamie McCarthy]

awilliam at whitemice.org writes:

> > adequate locking.  That would mean that each "peer" would have
> > to get an acknowledgment from each of the other peers that that
> > it has received a "record 53 of table mytable" is locked. 
> 
> Yes, and implement START/STOP operation notificaiton with
> timeouts, post-fault resyncronization, and multi-host rollback. 
> Yep.  And you wonder why almost no one support this? :)  Of
> course, almost no one needs it.
>
> You run the databases on high-end machines, most likely with
> fiber-channel interconnects. $$$,$$$.99 at a minumum.

MySQL Cluster does exactly this;  your application connects to a
single host machine, but only when all machines in the cluster
have commited the entire transaction does the host return the
status to your application.  Very tricky to get right, but they've
done it.

They do recommend special NICs designed for low latency, for
obvious reasons.  Not that expensive, maybe $1K a card, SCI is the
name I've heard.  But the DB machines can be low five figures
(each) or even high four figures... Opterons are cheap nowadays.

And Emic offers a very similar functionality, based on MySQL,
without requiring RAM residency.  <http://www.emicnetworks.com/>

> I know of an Informix cluster instance of two geographically
> distant servers interconnected by an OC-3.

Right, if your application is important enough that you need data
integrity and failover during a tornado, earthquake, etc., then a
geographically separated cluster is an expensive but viable
solution.
-- 
  Jamie McCarthy
 http://mccarthy.vg/
  jamie at mccarthy.vg