Aria: a fast and practical deterministic OLTP database

Abstract

© 2020, VLDB Endowment. Deterministic databases are able to efficiently run transactions across different replicas without coordination. However, existing state-of-the-art deterministic databases require that transaction read/write sets are known before execution, making such systems impractical in many OLTP applications. In this paper, we present Aria, a new distributed and deterministic OLTP database that does not have this limitation. The key idea behind Aria is that it first executes a batch of transactions against the same database snapshot in an execution phase, and then deterministically (without communication between replicas) chooses those that should commit to ensure serializability in a commit phase. We also propose a novel deterministic reordering mechanism that allows Aria to order transactions in a way that reduces the number of con icts. Our experiments on a cluster of eight nodes show that Aria outperforms systems with conventional nondeterministic concurrency control algorithms and the state-of-the-art deterministic databases by up to a factor of two on two popular benchmarks (YCSB and TPC-C).