A Delete-Aware Compaction Trigger Method for LSM-Tree Based Key-Value Stores

Abstract

LSM-tree based key-value stores show high performance in write-intensive workloads due to the out-of-place update structure. However, this structure requires additional storage space to store a lot of redundant data in database and results in high space amplification. In order to remove redundant data, most modern LSM-tree stores use a compaction triggered by the capacity in each level. But the size-based compaction trigger often occurs inadequate compaction frequency according to the workload, causing high write amplification and reduces overall performance. To address this, we introduce delete-aware compaction trigger which responds to the current workload’s deletion rates and reduces ineffi-cient compactions consisting of most valid records. We implemented Delete-Aware RocksDB on top of RocksDB, and we show that it outperforms by keeping low space amplification without write amplifica-tion cost compared to RocksDB. Furthermore, we show our system cooperated with Monkey, one of the state-of-the-art LSM-tree based key-value store, outperforms existing systems in terms of throughput.