Scalable Consistency for Data Replication

• Main Author: Fouto, Pedro Filipe Veiga
• Publication Date: 2024
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10362/179323
• topic: Data replication; Causal consistency; State machine replication; Edge computing; Geo-replication; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Distributed data storage solutions are key components of large-scale Internet services. The consistency guarantees provided by the protocols used to replicate data in such solu- tions can vary greatly, with some protocols providing strong guarantees while sacrificing availability, and others providing weaker guarantees while allowing higher availability. The choice of the replication protocol to use is often tied to the physical distribution of data replicas. As a general rule for Internet services, data replicated within a single datacen- ter is often replicated using protocols providing stronger consistency guarantees, while data replicated across multiple datacenters (i.e., geo-replicated) is often replicated using protocols providing weaker consistency guarantees. However, designing performant and fault-tolerant data replication solutions with data consistency that can scale in number of replicas is a challenging task. This is true not only for strong consistency models, but also for weaker consistency models, such as causal (and causal+) consistency. In this thesis, we propose to address the scalability and fault-tolerance limitations of providing consistency guarantees in data replication solutions, by addressing the entire spectrum of replication deployment scenarios, from single datacenter deployments, to geo-replicated, and edge computing deployments. To this end, we propose three main contributions, one for each of these deployment scenarios. First, we propose a new state-machine replication protocol, based on a new variant of Paxos, which improves on existing solutions by maximizing the throughput while avoiding performance degrada- tion when increasing the number of replicas. Second, we leverage the properties of the proposed protocol to replicate key components in geo-replicated causal consistency solu- tions, overcoming their fault-tolerance limitations while maximizing their performance. Third, by addressing the challenges of the edge computing environment, we propose a causal consistency data management solution that can efficiently scale to hundreds of edge locations.