Toward Continuous Service in Enterprise-Level Distributed Block Storage Systems

Abstract

This study investigates methods to reduce failure-induced downtime to the sub-second scale. We introduce a fully symmetric and decentralized architecture, eliminating single points of failure, together with an optimized failure detection mechanism and a temporary primary-replica based recovery procedure. The proposed design enables failure detection within 700 milliseconds and sustains uninterrupted I/O service by seamlessly promoting a secondary replica to a temporary primary, while concurrently restoring complete data to it and eventually transitioning it to a permanent role. We have implemented a Distributed Storage System, named ZDBS, in which the proposed mechanisms are fully integrated, and conducted extensive experimental evaluations. The results demonstrate that ZDBS substantially improves system availability across diverse failure scenarios, providing a robust solution for enterprise-level storage environments. Furthermore, we believe that the underlying principles of our approach can also be extended to other types of distributed systems to further enhance their availability.