Thanks, great article and very useful!

I wish I read an article like this half a year ago. I guess the approach I ended up using for my web-apps was a kind of horizontal partitioning but with all the nodes being independent and isolated (all in one). I have a dispatcher (a type of load balancer) whose only role is to send each user the sub-domain it needs to talk to. The dispatcher has its own database of username versus sub-domain. The dispatcher can if needed be automatically reconstructed from the various nodes. The nodes have each their own independent backup and recovery.