Group, majority, and strict agreement in timed asynchronous distributed systems

Abstract

Atomic broadcast is a group communication service that enables a team of distributed processes to keep replicated data 'consistent', despite concurrency, communication uncertainty, failures and recoveries. We investigate possible meanings for replicated data 'consistency' in timed asynchronous systems, subject to crash/performance process failures and omission/performance communication failures which may partition correct team members into isolated parallel groups. We propose three different replica consistency specifications: group agreement, majority agreement and strict agreement and give examples of atomic broadcast protocols that implement these specifications. The interface issues between the underlying membership services and the broadcast protocols that provide the above semantics are also addressed.