Chimera: Mitigating Ownership Transfers in Multi-Primary Shared-Storage Cloud-Native Databases

Download PDF

Abstract

Cloud-native database systems with multi-primary shared-storage architecture have emerged due to their superior performance over primary-secondary architecture on write-intensive workload scenarios. However, these systems face performance degradation as the proportion of shared data increases, adversely affecting their CostPerformance Ratio (CPR). In this paper, we identify frequent page ownership transfers between primaries as a key factor contributing to these performance bottlenecks. To address this challenge, we propose Chimera , a multi-primary database system that employs a two-phase transaction scheduling mechanism, combined with a delay-fetch ownership transfer strategy to effectively reduce the overhead of ownership transfers. Extensive experiments on SmallBank and TPC-C benchmarks demonstrate that Chimera outperforms existing schedule methods for multi-primary systems, achieving performance gains of 1.86 × ∼ 19.03 × on throughput.