Some data only makes sense when you know when it happened. Servers report CPU and memory usage every few seconds. Applications track request rate and latency. Sensors send temperature readings. Exchanges publish prices and trades. In all of these examples, the timestamp is not just another column. It is the main way we store, query, and eventually delete the data.

Time-series databases are built for this shape of workload. New points are added all the time, queries usually include a time range, and summaries over time windows matter more than fetching one record. Recent data is read far more often than old data. Old data often expires or is kept as coarser summaries.

A regular relational database can handle small or moderate time-series workloads if it has the right indexes and partitions. A dedicated time-series database becomes useful when the hard parts are no longer just storing rows, but handling high write volume, deciding how long data should live, compressing old data, keeping summaries, and controlling the number of unique series.

This chapter covers what makes time-series workloads special and how time-series databases are built to handle them.