I already have a ring of servers (message brokers, to be precise) that uses Cassandra to persist data needed to recover from a crash. We played with some configurations and realized that we get much better performance if we run Cassandra on the same box as the app. The row key is well distributed, so 1/n chance that the IO operation can be performed locally. The traffic is also well distributed (it's a safe assumption) so the question is, why rely on the row key to distribute the rows around the ring? Why can't I just assume the local shard is the master that I want to write to? This would give me a 100% chance that the IO operation can be performed locally.
I assume this is a bit of a corner case, so there's probably nothing pre-boxed for my situation. But is there something that would function as a good starting point?