Horizontally Scalable Implementation of a Distributed DBMS Delivering Causal Consistency via the Actor Model

Abstract

Causal Consistency has been proven to be the strongest type of consistency that can be achieved in a fault-tolerant, distributed system. This paper describes an implementation of D-Thespis, which is an approach that employs the actor mathematical model of concurrent computation to establish a distributed middleware that enforces causal consistency on a widely used relational database management system (RDBMS). D-Thespis prioritises developer experience by encapsulating the intricacies of causal consistency behind an interface that is accessible over standard REST protocol. Here, we discuss several novel results. Firstly, we define a method that builds a causally consistent DBMS supporting elastic horizontal scalability. Secondly, we deliver a cloud-native implementation of the middleware and provide results and insights on 6804 benchmark configurations executed on our implementation while running on a public cloud infrastructure across several data centres. The evaluation concerns transaction processing performance, an evaluation of our implementation’s update visibility latency, and a memory profiling exercise. The results of our evaluation show that under a transactional workload, a single-node installation of our implementation of D-Thespis is 1.5 times faster than a relational DBMS running serialisable transaction processing, while the performance of the middleware can improve by more than three times when scaled horizontally within the same data centre. Our study of the memory profile of the D-Thespis implementation shows that the system distributes its memory requirements evenly across all the available machines, as it is scaled horizontally. Finally, we also illustrate how our middleware propagates data changes across geographically-distributed infrastructures in a timely manner: our tests show that most of the effects of data change operations in one data centre are available in a remote data centre within less than 300 ms over and above the network round trip latency between the two data centres.