PromCon 2016

Talk abstract

The Prometheus Time Series Database

Speaker: Björn Rabenstein

Various time series databases (TSDBs) have been implemented on top of key-value stores with BigTable semantics. The TSDB that sits at the core of the Prometheus monitoring system started with a similar approach and was built on top of LevelDB. The Prometheus server as we know it today, however, uses a highly optimized custom storage layer for bulk sample data, enabling a single server to sustain an ingestion rate of 500,000 samples per second belonging to millions of time series. Very recent improvements of the in-memory representation of sample data resulted in an outstanding compression level of 1.3 bytes per sample in a typical production setup. A journey from fundamental challenges of TSDB design to details of the Prometheus storage layer.