Year: 2024
Type: dissertation
Abstract: Replication has been a common strategy for improving the dependability of database systems. However, the use of database systems in demanding environments and critical applications calls for the enforcement of stronger dependability properties, in face of generalized fault models that include software bugs and malicious behavior. Commercial database systems assume a crash-recovery model, typically favoring a primary-backup approach for performance and scalability, which is unsuitable for Byzantine fault tolerance as a faulty primary replica can simply propagate errors to the backup replicas. Meanwhile, proposals for Byzantine fault-tolerant replication protocols can mostly be found in the literature and academic tools. However, with very few exceptions, proposed replication protocols, even for Byzantine fault-tolerance, assume replicas are homogeneous and, at the same time, that faults are independent. We question this assumption, particularly considering the complexity of software related to databases, addressing it with a study of the diversity and known vulnerabilities in open-source database software. Experience with existing systems, as captured by vulnerability collections, exposes the possible impact on homogeneous replication. Our study also shows there is sufficient diversity for heterogeneous replication to be feasible, allowing the use replicas with different database engines, software stacks, and even data models. We present data that demonstrates current solutions are suboptimal and propose a novel architecture that leverages the capabilities of LLMs in code translation to enable heterogeneous Byzantine fault-tolerant replication among replicas with multiple data models. We show how LLMs can be useful with practical examples and perform an evaluation of the final system, highlighting a trade-off between fault tolerance and performance.
Author: Imperadeiro, João Ribeiro
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Universidade de Coimbra
OpenAlex ID: W7138825684