Time series data

Time series data means readings stamped with the time they were taken, such as a sensor measured every second or a shop's sales added up every day. Its main feature is that the time is part of every reading, and the order of readings matters. A series of readings can show a trend, a repeating pattern or a sudden change.

Here is how it works in general. A source, such as a sensor or an app, produces a reading at regular times. The system stores each reading together with its time. Because the readings arrive constantly and are almost never changed, special databases are used that can store millions of readings a second and answer questions such as "what was the average between 2 pm and 3 pm". Old readings are often summarised, for example from one reading a second to one a minute, to save space.

An ordinary database struggles with this volume. A plant with 10,000 sensors that each give one reading a second produces 864 million readings a day. Readings may also arrive late or out of order, because the network drops messages. So the database stores each reading with the time it was measured, not the time it arrived. Questions are usually asked over a window of time: the average per hour, the highest value per day, or how fast a value is rising. Old readings are summarised and then deleted after a set period.

Take, for example, a smart electricity meter that records the household's use every 15 minutes. The utility uses the series to bill by time of day and to plan supply; see meter data networks. Consider a cloud company that records the load on each of its servers every second and raises an alert when it jumps; see telemetry. Say a shop chain records sales per store per day. It can then forecast the next festival season using demand forecasting.

Time series data is the raw material for predictive maintenance and for anomaly detection. The readings often come from industrial IoT, and a plant keeps them in a plant historian.

How is it different from an ordinary table of records? In an ordinary table, such as a list of customers, a row can be changed at any time and time may not matter. In a time series, rows are added, never changed, and the time order is the whole point.