Planning a Large RDBMS

Hi folks,

We've been using MongoDB for the past couple years and are talking about
moving to a more traditional (i.e. stable) database system for OLTP. I've
been using RDBMS for OLAP already, but the demands are somewhat different.
But we figure, if Facebook can do it, we can. The expertise around here is
more suited for MySQL, but I, for one, am fond of the features in Postgres
that you can't find in MySQL. Our Mongo system is sharded across 11 logical
nodes on Amazon's EC2. Our database is in the hundreds of gigabytes (i.e.
too big for memory on one machine) and we can expect it to exceed
a terabyte before too long. We have the distinct advantage of being able to
plan this from the ground up. I've read a bit about partitioning and table
spaces, but from what I can tell, those solutions still require the
database to be hosted by a single machine which may not suffice for our
purposes.

I realize this is a big question, and possibly too big to answer here. But
I would be grateful even for some website or chapter recommendations.

Thank you in advance,
-Alessandro