Fast transactions in distributed and highly available databases

Abstract

Many modern data-oriented applications are built on top of distributed OLTP databases for both scalability and high availability. However, when running transactions that span several partitions of the database, signicant performance degradation is observed in existing distributed OLTP databases. In this thesis, we develop three systems -- (1) STAR, (2) COCO, and (3) Aria -- to address the inefficiency and limitations of existing distributed OLTP databases while using dierent mechanisms and bearing various tradeoffs. STAR eliminates two-phase commit and network communication through asymmetric replication. COCO eliminates two-phase commit and reduces the cost of replication through epoch-based commit and replication. Aria eliminates two-phase commit and the cost of replication through deterministic execution. Our experiments on two popular benchmarks (YCSB and TPC-C) show that these three systems outperform conventional designs by a large margin. We also characterize the tradeoffs in these systems and the settings in which they are most appropriate.