Perpetual Computing Cluster

Consider a highly fault tolerant application clustered across N servers from a pool of T servers, with N << T. Where the cluster continuously loses, destroys, and instantiates nodes, while replicating or balancing state across nodes.

Assuming geographic or network location was transparent to the application, this cluster could effective manage itself continuously over different hardware on different networks in different geographies making it as resilient as the total system of servers hosted by the provider. Distribution across providers would be even more interesting.

This all is effectively the same as a P2P network running on various PC around the world hosting files or computing PI. But would be more reasonable for business to adopt since it would be intended for their business goals, not the entertainment of college kids and researchers.

Of course one of the most obvious systems that could run on this sort of cluster would be a Hadoop distributed file system.

Or even S3 itself.