Causal consistency through a novel distributed middleware over strongly consistent transaction processing

Abstract

Our research deals with the concept of causal consistency of data in the context

of transactional information systems with scalability and high availability require- ments. We deal with the consistency of data which is stored and replicated in

multiple physical locations. Given the data store’s distributed nature, a new set of data inconsistency issues arise. These cause clients to get an inconsistent, and therefore possibly incorrect, view of the data, yielding application errors and even susceptibility to security vulnerabilities. Most problems do not impact centralised databases, but centralised databases do not provide the resiliency and performance characteristics required by modern enterprise transactional information systems. We focus on this set of data inconsistency problems, and propose solutions to strengthen consistency guarantees without jeopardising the benefits of a distributed database. We model causal consistency, the strongest type of consistency that can

be implemented in fault-tolerant, scalable databases, using the Actor model of com- putation. The model is then implemented on top of commercially-ready relational

database management systems that are built to provide strong consistency. Data Inconsistency, Transaction Inconsistency and Integrity Invariant Violation

are three related, but distinct, problems tackled in this research. For each prob- lem, we review the literature as well as design, implement and evaluate a novel

solution. Our work shows that it is possible to have a distributed middleware that implements causal consistency with transaction consistency and integrity invariant

preservation over a set of disconnected relational databases deployed within geo- graphically distributed data centres. Thus, our approach addresses each problem

whilst answering to the scalability and resiliency needs of modern systems. Empirical results show that our middleware achieves better performance when compared to a single-node (i.e., non-distributed) relational database management

system. We also extend our solution for Data Inconsistency and deploy the middle- ware on many machines within a data centre. In doing so, we identify and propose

solutions for the complexities that arise from scaling the middleware horizontally, whilst our benchmarks show a significant increase in the amount of operations that can be processed at each data centre, and that data changes are replicated across geographically distributed instances of the system within acceptable timeframes.