DISAGGREGATED IN-MEMORY DATABASE SYSTEMS

Memory disaggregation has emerged as a promising architectural paradigm for modern data centers, offering elastic memory provisioning, improved resource utilization, and reduced total cost of ownership. By decoupling compute and memory resources and connecting them via high-performance RDMA networks, disaggregated memory enables in-memory database systems to scale beyond the limitations of monolithic servers. However, this architectural shift also introduces fundamental challenges in efficient data access, synchronization, transaction processing, and fault tolerance.This dissertation investigates how to design high-performance and reliable in-memory database systems over RDMA-enabled disaggregated memory. It argues that traditional database architectures, which assume tightly coupled compute and memory, are ill-suited for this environment, and that architecture-native system redesign is required. To this end, the dissertation presents three complementary works that collectively address the core challenges of indexing, cache coherence, and transaction management in disaggregated settings.First, the dissertation studies index structures over remote memory and introduces an LSM-based indexing approach tailored to disaggregated architectures, enabling efficient data organization and scalable write performance under remote access constraints. Second, it introduces a cache-coherence abstraction that provides correct and efficient shared-data access using one-sided RDMA operations, serving as a general foundation for building database components over disaggregated memory. Third, building on this abstraction, the dissertation develops a multi-primary storage engine that supports multi-version concurrency control and high availability while minimizing remote communication overhead.Together, these contributions demonstrate that with careful, architecture-aware design, in-memory database systems can achieve high performance, scalability, and reliability over disaggregated memory. Beyond specific system designs, this dissertation provides broader insights into building data-intensive systems on emerging disaggregated architectures.