While doing the demo for JUDCon, I realised that smart routing on the client side always choses the same node for the same key.

IOW, assume you have k1 stored in nodes A, B and the cluster is formed of nodes A, B and C. Now, when the client figures out that the owner is A, any further requests on k1 will be sent to A. This has the possibility of overloading A.

I think the client should, from the given list of owners of a key, do round robin between them for requests on the same key.