Achieving strongly consistent and efficient geo-distributed transactions in edge computing

Abstract

The prosperity of edge computing fosters the deployment of distributed storage systems across many near-client and edge data centers to provide low and stable data access latency for end-users and edge computing applications. Many emerging real-world edge computing applications host mission-critical services (e.g., smart-city and vehicular-network management systems, UAV and flight airspace scheduling systems, and edge-based financial traditional systems) which desire strong consistency and stringent client-perceived latency. Moreover, since edge data centers are usually owned by different parties with less powerful security defense systems, tolerating malicious participants and maintaining high robustness are also essential for a distributed storage system deployed across edge data centers.

Unfortunately, our study found that existing distributed storage systems that are designed to be deployed across a few core data centers cannot meet the efficiency and robustness requirements if they are deployed across edge or near-client data centers. On the other hand, existing distributed storage systems for edge computing applications are not designed to provide strong consistency, making their guarantees insufficient for the emerging mission-critical applications on the edge.

This thesis showcases the design, implementation, and evaluation of three efficient, robust, and scalable data-intensive storage systems with strong consistency guarantees for mission-critical edge computing applications. The thesis first presents EGES, an efficient and robust (DoS-resistant) consensus protocol among many nodes in edge computing with the linearizability consistency guarantee, which works as a basic building block for the data storage layer. The second system DAST adds sharding capability and ACID guarantees to data access requests and solves the blocking problem introduced when preserving transactional isolation, making the data storage system an efficient and scalable OLTP (OnLine Transactional Processing) database. The third work first theoretically proves an important impossibility result in HTAP (Hybrid Transactional and Analytical Processing) system design: one must tradeoff between the strongest consistency level and performance scalability. Guided by this result, we design PRIST, which adds real-time query processing capability to the data storage system, making it an efficient and scalable HTAP system with the strongest consistency guarantee to date.