Time-series databases

A time-series database is built for data that arrives as a stream of timestamped measurements, like server CPU every ten seconds, sensor readings or stock prices. Because new points almost always arrive in time order and old points rarely change, it can append them quickly and store them very compactly, compressing the small differences between neighboring values. Queries are usually about ranges and trends, such as “average latency per minute over the last day”, so it’s organized around time windows. To keep storage under control, it downsamples older data, keeping just the min, max and average per hour or day, and deletes raw points past a retention period.