A Semantic Consistency Model to Reduce Coordination in Replicated Systems

• Autor(a) principal: Gomes, Nuno Filipe Estêvão
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/127080
• topic: Distributed Systems; replication; synchronization; consistency; concurrency; invariants; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Large-scale distributed applications need to be available and responsive to satisfy millions of users, which can be achieved by having data geo-replicated in multiple replicas. However, a partitioned system cannot sustain availability and consistency at fully. The usage of weak consistency models might lead to data integrity violations, triggered by problematic concurrent updates, such as selling twice the last ticket on a flight company service. To overcome possible conflicts, programmers might opt to apply strong consistency, which guarantees a total order between operations, while preserving data integrity. Nevertheless, the illusion of being a non-replicated system affects its availability. In contrast, weaker notions might be used, such as eventual consistency, that boosts responsiveness, as operations are executed directly at the source replica and their effects are propagated to remote replicas in the background. However, this approach might put data integrity at risk. Current protocols that preserve invariants rely on, at least, causal consistency, a consistency model that maintains causal dependencies between operations. In this dissertation, we propose a protocol that includes a semantic consistency model. This consistency model stands between eventual consistency and causal consistency. We guarantee better performance comparing with causal consistency, and ensure data integrity. Through semantic analysis, relying on the static analysis tool CISE3, we manage to limit the maximum number of dependencies that each operation will have. To support the protocol, we developed a communication algorithm in a cluster. Additionally, we present an architecture that uses Akka, an actor-based middleware in which actors communicate by exchanging messages. This architecture adopts the publish/subscribe pattern and includes data persistence. We also consider the stability of operations, as well as a dynamic cluster environment, ensuring the convergence of the replicated state. Finally, we perform an experimental evaluation regarding the performance of the algorithm using standard case studies. The evaluation confirms that by relying on semantic analysis, the system requires less coordination between the replicas than causal consistency, ensuring data integrity.