What is a time series database (TSDB), and why is it faster than a relational database for metrics?

Its root structure is a timeline rather than tables of rows: every value is attached to a point in time, so "give me metric X from t1 to t2" is the native query and needs no transposing.

In a relational database, a time range has to be assembled: find the rows, sort them by timestamp, turn them into a series. In a TSDB the data already is a series. Each metric is a sequence of (timestamp, value) pairs, stored and compressed in time order. Asking for a range is a sequential read of exactly the stretch you want.

Because the query matches the storage layout, the transposition step disappears, and with it the performance bottleneck. That makes a TSDB the right store for trending, and with a rule engine on top it can do monitoring too.

Examples: Prometheus, InfluxDB, Graphite, OpenTSDB.

Go deeper:

• Time series database — Wikipedia — definition, typical features and the main products.