Anti-Entropy Bandits for Geo-Replicated Consistency

Eventually consistent systems can be made more consistent by reducing the time until a write is fully replicated, thereby improving global update visibility. While gossip-based anti-entropy methods scale well, random selection of anti-entropy partners is less than efficient. Moreover, while eventual consistency may be consistent enough in a single data center, geographic replication increases visibility latency and leads to externally observable inconsistencies. In this paper, we explore an improvement to pairwise, bilateral anti-entropy; instead of uniform random selection, we introduce reinforcement learning mechanisms to assign selection probabilities to replicas most likely to have information. The result is more efficient replication, faster visibility, and stronger eventual consistency while maintaining high availability and partition tolerance.