Lumière: Byzantine View Synchronization

Many distributed protocols in the partial synchrony setting with Byzantine nodes divide the local state of the nodes into views, and the transition from one view to the next dictates a leader change. In order to provide liveness, all honest nodes need to stay in the same view for a sufficiently long time. This requires intricate mechanisms that are typically intertwined with the rest of the protocol, making it hard to understand and to reason about. Furthermore, state-machine replication, which is consisted of multiple instances of single-shot consensus, can use the same view synchronization protocol. We define the Byzantine View Synchronization problem, which is responsible for eventually bringing all nodes to the same view for a sufficiently long time. Two approaches for implementing a protocol that achieves view synchronization exhibit the following tradeoffs: a view doubling solution has zero communication costs but unbounded latency, while a broadcastbased solution has quadratic communication costs but constant latency. We describe both protocols, prove their correctness, and also introduce a third protocol, named Lumi\`ere, that has optimistically linear communication complexity and constant latency, and faced with benign failures, has expected linear communication and constant latency. Lumi\`ere is particularly useful for a family of consensus protocols that exhibit linear communication under various circumstances.