Wow, this looks like it could yield some impressive improvements over the naive replication strategies in Hadoop today.