Dostoevsky: Better Space-Time Trade-Offs for LSM-Tree Based Key-Value Stores via Adaptive Removal of Superfluous Merging

In this paper, we show that all mainstream LSM-tree based key-value stores in the literature and in industry are suboptimal with respect to how they trade off among the I/O costs of updates, point lookups, range lookups, as well as the cost of storage, measured as space-amplification. The reason is that they perform expensive merge operations in order to (1) bound the number of runs that a lookup has to probe, and to (2) remove obsolete entries to reclaim space. However, most of these merge operations reduce point lookup cost, long range lookup cost, and space-amplification by a negligible amount. To address this problem, we expand the LSM-tree design space with Lazy Leveling, a new design that prohibits merge operations at all levels of LSM-tree but the largest. We show that Lazy Leveling improves the worst-case cost complexity of updates while maintaining the same bounds on point lookup cost, long range lookup cost, and space-amplification. To be able to navigate between Lazy Leveling and other designs, we make the LSM-tree design space fluid by introducing Fluid LSM-tree, a generalization of LSM-tree that can be parameterized to assume all existing LSM-tree designs. We show how to fluidly transition from Lazy Leveling to (1) designs that are more optimized for updates by merging less at the largest level, and (2) designs that are more optimized for small range lookups by merging more at all other levels. We put everything together to design Dostoevsky, a key-value store that navigates the entire Fluid LSM-tree design space based on the application workload and hardware to maximize throughput using a novel closed-form performance model. We implemented Dostoevsky on top of RocksDB, and we show that it strictly dominates state-of-the-art LSM-tree based key-value stores in terms of performance and space-amplification.