Couple of interesting posts on Twitter scaling issues.
If that’s the case I wonder what an Erlangish solution would be. There’s no real reason for the Twitter *database* to be centralized at all. (In this sense, Winer’s thoughts on total decentralization are apt.) Could you not have multiple nodes, each with it’s own database of some users. These nodes are going to have to talk to each other only when someone pulls a page out.
You could also, presumably, in the extreme case have entirely separate servers for different users, and only integrate the overview of all people being followed “at the glass” in the browser or RIA client.