BTrDB: Optimizing Storage System Design for Timeseries Processing

The increase in high-precision, high-sample-rate telemetry timeseries poses a problem for existing timeseries databases which can neither cope with the throughput demands of these streams nor provide the necessary primitives for effective analysis of them. We present a novel abstraction for telemetry timeseries data and a data structure for providing this abstraction: a time-partitioning version-annotated copy-on-write tree. An implementation in Go is shown to outperform existing solutions, demonstrating a throughput of 53 million inserted values per second and 119 million queried values per second on a four-node cluster. The system achieves a 2.9× compression ratio and satisfies statistical queries spanning a year of data in under 200ms, as demonstrated on a year-long production deployment storing 2.1 trillion data points. The principles and design of this database are generally applicable to a large variety of timeseries types and represent a significant advance in the development of technology for the Internet of Things.