Bruno,
> Which brings up another question: why not just cluster at the hardware
> layer? Get an external fiberchannel array, and cluster a bunch of dual
> Opterons, all sharing that storage. In that sense you would be getting
> one big PostgreSQL 'image' running across all of the servers.
>
> Or is that idea too 90's? ;-)
No, it just doesn't work. Multiple postmasters can't share one database.
LinuxLabs (as I've gathered) tried to go one better by using a tool that
allows shared memory to bridge multple networked servers -- in other words,
one postmaster controlling 4 or 5 servers. The problem is that IPC via this
method is about 1,000 times slower than IPC on a single machine, wiping out
all of the scalability gains from having the cluster in the first place.
--
--Josh
Josh Berkus
Aglio Database Solutions
San Francisco