Constructing and analyzing the LSM compaction design space

Log-structured merge (LSM) trees offer efficient ingestion by appending incoming data, and thus, are widely used as the storage layer of production NoSQL data stores. To enable competitive read performance, LSM-trees periodically re-organize to form a tree with levels exponentially increasing capacity, through iterative compactions. Compactions fundamentally influence performance an LSM-engine in terms write amplification, throughput, point range lookup space delete performance. Hence, choosing appropriate compaction strategy is crucial and, at same time, hard LSM-compaction design vast, largely unexplored, has not been formally defined literature. As result, most LSM-based engines use fixed strategy, typically hand-picked engineer, which decides how when compact data. In this paper, we present LSM-compactions, evaluate state-of-the-art strategies respect key metrics. Toward goal, our first contribution introduce set four primitives that can define any strategy: (i) trigger, (ii) layout, (iii) granularity, (iv) movement policy. Together, these synthesize both existing completely new strategies. Our second experimentally analyze 10 We 12 observations 7 high-level takeaway messages, show LSM systems navigate space.