TSCache
Jian Liu,Kefei Wang,Feng Chen
DOI: https://doi.org/10.14778/3484224.3484225
IF: 2.5
2021-01-01
Proceedings of the VLDB Endowment
Abstract:Time-series databases are becoming an indispensable component in today's data centers. In order to manage the rapidly growing time-series data, we need an effective and efficient system solution to handle the huge traffic of time-series data queries. A promising solution is to deploy a high-speed, large-capacity cache system to relieve the burden on the backend time-series databases and accelerate query processing. However, time-series data is drastically different from other traditional data workloads, bringing both challenges and opportunities. In this paper, we present a flash-based cache system design for time-series data, called TSCache . By exploiting the unique properties of time-series data, we have developed a set of optimization schemes, such as a slab-based data management, a two-layered data indexing structure, an adaptive time-aware caching policy, and a low-cost compaction process. We have implemented a prototype based on Twitter's Fatcache. Our experimental results show that TSCache can significantly improve client query performance, effectively increasing the bandwidth by a factor of up to 6.7 and reducing the latency by up to 84.2%.