Using NFS doesn't seem like a good idea for this situation - from the operations perspective, NFS in production is going to be hell anyway. A better way would

Using NFS doesn't seem like a good idea for this situation - from the operations perspective, NFS in production is going to be hell anyway. A better way would be (since Tokyo Cabinet is thread safe) having a separate thread/process which receives messages (using a message queuing system) that update the database files (ala MySQL replications). (Or you can separate out the db layer as a separate network service and partition the writes/reads to go to a specific several servers based on the key. Tokyo Cabinet even includes a memcached API compatible network layer for this - the Tokyo Tyrant).