You'll have to design your stuff to tolerate failure; use things like Chef for quick deployments so that provisioning new servers isn't a huge pain in the ass.

You'll have to design your stuff to tolerate failure; use things like Chef for quick deployments so that provisioning new servers isn't a huge pain in the ass. Further, the biggest thing to keep in mind is that your resources are shared. This is not too big of a deal for some applications, but for a database server, this can spell certain death (I/O doesn't virtualize too well; you can mitigate some of this by deploying a lot of slave machines for your reads.)