A Structural Classification of Integrated Replica Control Mechanisms

We classify practical replica control (RC) mechanisms according to two main criteria: (1) the abstract RC protocol, e.g., how many physical copies are written for a logical update, and (2) the kind of integration between the RC mechanism and system services such as atomic broadcast, commit protocols, and concurrency control. We describe about a dozen representative RC mechanisms (most of them implemented) classified by their integration with communications, concurrency control, or both. This classification based on a few essential components of RC mechanisms show that practical replicated systems overwhelmingly adopt the read-one/write-one RC protocol (read/write one physical copy for each logical read/write), due to its low overhead and high availability, despite some inherent inconsistency. Furthermore, this trend points to the need for the integration of weak consistency methods (e.g., epsilon serializability) with read-one/write-one RC protocols if replicated databases are to be realized in practice.