HTAP-friendly Version Management Architectures for MVCC Databases

Abstract

Multi-version concurrency control (MVCC) and design principles thereof are ingrained in MVCC databases for decades. It facilitates scalable transaction processing by providing point-in-time consistency, rescuing conflict operations that may otherwise be rejected in non-MVCC systems. Despite its favorable capability, contemporary MVCC systems are still vulnerable to the age-old problem: long-lived transactions. The presence of long-lived transactions impacts the stability of the multi-version nature, decreasing all aspects of performance metrics. Furthermore, the emergence of hybrid transactional/analytical processing (HTAP) aggravates the situation in that the complex analytic queries likely trigger the long-lived transactions while online transactions intensively expand the version space. This thesis presents a series of three version management architectures, vDriver, vWeaver, and DIVA. These architectures are designed by different approaches—from a practical remedy preserving the prevailing architecture to a comprehensive redesign respecting the separation of concerns principle—but share the same goal of breaking through the space-time tradeoff of conventional MVCC systems. To validate the proposed designs, we integrated each version management architecture into PostgreSQL and MySQL, well-known full-fledged MVCC database systems. The experimental evaluations demonstrate that the proposed designs generally improve all aspects of performance metrics, making MVCC databases HTAP-friendly.