In Search of a Scalable Raft-based Replication Architecture

Providing a consistent replicated log across different servers, the Raft consensus protocol greatly facilitates the design of fault-tolerant services. However, due to the protocol following the principle of a single strong leader, such architectures in general do not scale with the number of cores and/or network cards available on each server. To address this problem, we present the Niagara replication architecture, which makes it possible to build scalable systems while still relying on Raft for consensus. In particular, we show how Niagara parallelizes the process of appending new log entries across multiple Raft instances and discuss Niagara's support for read operations with different consistency requirements.