PhD Defense: High Performance Distributed Transactions for Multi-region Database Systems

https://umd.zoom.us/my/ctring?pwd=NGFkcWFtU2VFcnl3c253RzhmUjI2QT09

The increased popularity of global applications has spurred the demand for a more advanced database system layer in multi-region cloud infrastructures. This layer should seamlessly abstract the distributed nature of the system behind a strong data consistency interface, enabling developers to focus on application logic without delving into complex data race issues. Simultaneously, it needs to exhibit resilience in the face of calamities, ranging from hardware failures to natural disasters that shut down entire data centers. However, these goals are usually at odds with the need for high performance and scalability. Guaranteeing strong data consistency model necessitates additional coordination rounds between nodes to ensure the proper ordering of read and write operations. Enhancing availability and durability necessitates data replication across multiple locations. These measures, coupled with the potential for high-latency cross-region communication, can significantly degrade performance. Consequently, existing solutions typically either compromise on weaker consistency levels or incur performance and scalability penalties to achieve better consistency guarantees.This dissertation addresses these challenges in three directions. First, we adopt a retrospective approach by enhancing an existing production-ready shared-storage architecture, derived from the Amazon Aurora database system. Our proposed modifications enable Aurora-style systems to support multiple writer nodes across geographically dispersed regions, significantly reducing latency and eliminating scaling bottlenecks. Second, we adopt a forward-looking perspective by investigating a more recently proposed architecture known as deterministic database systems. We develop a new protocol within this architecture that facilitates the processing of strictly serializable multi-region transactions with minimal performance degradation under high-contention workloads, achieving throughput improvements by an order of magnitude relative to state-of-the-art approaches and reducing latency by up to a factor of five. Finally, we take a comprehensive view by performing an empirical study of various real-world applications, reassessing the assumptions of recent database system proposals, and offering valuable insights to inform future research in transactional database systems.