Ambulance: saving BFT through racing

2026-06-23

AI Summary

Existing practical Byzantine Fault Tolerance (BFT) protocols rely on timeout mechanisms to detect leader failures, often striking an unstable balance between aggressive and conservative strategies, which degrades performance. This work proposes Ambulance, a novel BFT protocol that introduces an “in-protocol racing” mechanism, eliminating reliance on timeouts by enabling replicas to naturally compete for leadership through parallel execution of protocol steps. Coupled with a collaborative asynchronous recovery scheme, Ambulance achieves timeout-free failure detection and leader transition. The design maintains high throughput and low latency under normal operation while rapidly responding to slow-leader failures, thereby achieving both high performance and strong robustness, significantly outperforming conventional timeout-based BFT protocols.

Abstract

Today's practical Byzantine Fault Tolerant (BFT) state machine replication deployments are vulnerable to slowdowns. The main culprit is timeouts. Aggressive timeouts spuriously trigger expensive leader changes, while conservative timeouts leave the system idle and let slowdowns severely inflate latency. Two main alternatives exist: hedging, which improves recovery from slow leaders but still incurs a time-based hedging delay, and cooperative asynchronous protocols, which recover quickly from slowdowns but suffer from high common-case latency and low throughput. This paper presents Ambulance: a BFT state machine replication protocol that sidesteps this trade-off through protocol-rigged races, where replicas, rather than race against the clock, race against each other by executing protocol steps. This enables Ambulance to achieve high throughput and low latency comparable to state-of-the-art timeout-based BFT, while matching the robustness of cooperative approaches.

Problem

• Byzantine Fault Tolerance
• state machine replication
• timeout
• latency
• throughput

Innovation

• Byzantine Fault Tolerance
• state machine replication
• protocol-rigged races
• leader election
• timeout-free consensus