mutant

Abstract

Conventional wisdom in database systems says there is a trade-off between database storage cost and the access latency: you can build a low-cost, high-latency database, or a low-latency, high-cost database, but not both. We present Mutant, a layer for LSM-tree cloud data stores that balances this trade-off. The key observation is that high temporal locality of web workloads coupled with sequential batching of LSM-tree tablets means that colder data lives on older tablets, allowing them to be moved to colder storage. In addition to controlling the cost and access latency trade-off, our implementation of Mutant on top of RocksDB shows that Mutant reduces cloud storage costs up to 90% when access latencies are kept fixed, and decreases latencies by 79% when cloud storage costs are fixed.