A new consensus protocol named A-greement has been discovered with the assistance of artificial intelligence, promising substantial improvements in latency for geo-replicated state machines. The research, detailed in a paper on arXiv, shows that A-greement reduces mean latency in key-value store operations by up to 33.3% compared to existing leading protocols.
Consensus protocols are fundamental to distributed computing, ensuring that multiple systems can agree on a single state or value, even in the presence of failures. They are critical for state machine replication (SMR), a technique used to build fault-tolerant systems by replicating data across multiple servers. Despite decades of research, finding low-latency solutions has remained a challenge, requiring a blend of protocol knowledge, rigorous reasoning, implementation, and testing.
The AI-driven discovery process involved a system called AI-greement. This system integrates a knowledge base of protocol concepts, a modular implementation framework, and a verification system using the Lean theorem prover. AI-greement was used to autonomously discover, implement, experiment with, and verify candidate protocols. This approach not only identified A-greement but also uncovered previously unknown vulnerabilities in existing consensus protocols.
A-greement achieves its performance gains through a combination of techniques. These include lease-authorized local reads, direct client learning for leader-ordered writes, and a fast, fixed-quorum per-key write path. The specific path chosen depends on the network topology and workload. In tests across six deployments with either five or nine replicas, A-greement reduced mean latency by 12.2% on write-heavy workloads and 33.3% on read-heavy workloads. For read-only workloads, its performance matched the fastest baseline protocols.
The researchers emphasize that while AI has proven effective in accelerating discovery, human experts remain essential. They suggest that AI agents are reshaping the role of researchers rather than replacing them, aiding in tasks like implementation and verification, but still requiring human insight for direction and interpretation.