Most, if not all, modern network controllers acknowledge that they are distributed systems and must deal with how to replicate state. There is a trade-off between strong consistency (with easier debugging, programming models, and understandability) on the one hand and tolerance for high latency between controllers and possible partitions on the other hand. To address this, we have extended a strongly-consistent data store (based on RAFT) to also allow for best-effort replication to additional nodes. This allows there to be a small number of nodes (with low latency between them) providing a strongly-consistent data store, while efficiently replicating state to other entities including other controllers acting as warm standbys to handle disaster recovery. Future extensions can allow simple federation by leveraging selective best-effort replication to exchange state without tight coupling.
