Replica Divergence in Data-Centric Consistency Models

摘要

Eventual consistency is demanded nowadays in highly scalable and available geo-replicated services. According to the CAP theorem, when network partitions may arise a distributed service should choose between being strongly consistent or being highly available. Since scalable services must be available, relaxed consistency is the regular choice. Eventual consistency is not a regular data-centric consistency model, but only a state convergence property to be added to a relaxed consistency model. This paper discusses which data-centric consistency models are not implicitly convergent and, because of this, provide an adequate basis for building eventually consistent services.