Rethinking Update-in-Place Key-Value Stores for Modern Storage

Abstract

Several widely-used key-value stores, like RocksDB, are designed around log-structured merge trees (LSMs). Optimizing for the performance characteristics of HDDs, LSMs provide good write performance by emphasizing sequential access to storage. However, this approach negatively impacts read performance: LSMs must employ expensive compaction jobs and memory-consuming Bloom filters in order to achieve reasonably fast reads. In the era of NVMe SSDs, we argue that this trade-off between read performance and write performance is sub-optimal. With enough parallelism, modern storage media have comparable random and sequential access performance, making update-in-place designs, which traditionally provide high read performance, a viable alternative to LSMs.

In this thesis, based on a research paper currently under submission, we close the gap between log-structured and update-in-place designs on modern SSDs by taking advantage of data and workload patterns. Specifically, we explore three key ideas: (A) record caching for efficient point operations, (B) page grouping for high-performance range scans, and (C) insert forecasting to reduce the reorganization costs of accommodating new records. We evaluate these ideas by implementing them in a prototype update-in-place key-value store called TreeLine. On YCSB, we find that TreeLine outperforms RocksDB and LeanStore by 2.18× and 2.05× respectively on average across the point workloads, and by up to 10.87× and 7.78× overall.