Abstract:
Distributed systems have traditionally favored responsiveness over consistency. For example, routing protocols apply received updates immediately even though such responsiveness comes at the cost of routing loops and blackholes, middleboxes manage and update routing state without making the updates durable, and distributed storage systems trade off consistency to achieve greater availability.

Our position is that consistent state in a distributed system not only makes its behavior more predictable, but surprisingly also improves its availability and performance. We will illustrate this using three case studies in this talk: Consensus Routing -- a consistency-first approach that cleanly separates safety and liveness, ETTM -- a new scalable and fault tolerant network manager that enables participating end-hosts to cooperatively manage network resources. and Scatter -- a scalable and distributed key-value storage system that provides linearizable consistency even under adverse circumstances.