OpenDaylight Summit 2014

The CAP theorem states that any distributed system, e.g., an OpenDaylight controller cluster, can have at most two of consistency, availability and partition-tolerance. As preventing partitions is hard, most distributed systems trade off between consistency and availability. SDN is a particularly challenging area for CAP trade-offs as partitions are likely to have clients, e.g., switches, on both sides. Thus, to keep the controller cluster available for all switches, it must seemingly give up consistency. Despite this, I believe we can do better than just dealing with eventual consistency. This talk will present these trade-offs and present several approaches that can provide minimal, and predictable, inconsistency while remaining available. This will include, exploiting data locality and a model for forking consistency.