Abstract

Multi-Raft architectures partition state machines into multiple distinct consensus groups to achieve horizontal scale. However, asymmetric network partitions and transient storage stalls can trigger cascading leader elections. This paper presents an empirical evaluation of latency under simulated adversarial partitions in multi-region deployments.

Research Findings

• Tail latencies under single-node partition regimes.
• Pre-vote algorithms drastically cut unneeded term bumps.
• Optimizing heartbeat intervals versus false election triggers.