Empirical study on a RocksDB multi-threaded compaction

Abstract

RocksDB is a high-performance key-value storage engine which offers a variety of features. Applications may adjust options to use the key-value store optimized for their purpose. RocksDB needs a background job called compaction. Compaction is an operation to remove duplicate or unnecessary records and improve lookup performance. If background jobs are delayed for too long, the latency of the user request may increased rapidly. This is called a latency spike. Many studies have been proposed so far to improve databases by changing the internal structure. However, these studies significantly changed the internal structure, resulting in high implementation cost and incompatible with existing databases. In addition, designs for special hardware are difficult to use universally. In this work, we analyze the operation and effect of subcompaction. Subcompaction is one of the features of RocksDB. Subcompaction improves compaction by distributing the compaction process using multiple threads. According to our observations, we found that the subcompaction reduces compaction latency by an average of 49\% and increased the throughput by up to 30\%. In addition, subcompaction mitigates the performance bottleneck caused by the slow background jobs by utilizing the CPU 18\%p more on average. Subcompaction is the original feature of RocksDB that improves background jobs at a low cost without compatibility and universality issues.