Time series

EngineersOperationsAnalysts

In one minute

A time series is one signal, measured repeatedly, a stream of timestamped values from a single instrument, or from a function that computes one. Connected series are the senses of a digital twin, the layer that keeps the model in step with the plant.

Two decisions matter when you create one: the unit and the value type. The value type is fixed forever at creation, so it is worth thirty seconds of thought.

Creating a series

Name it after the measurement

One series per measurement. Name it after the instrument loop where one exists: 21-PT-1234 Separator inlet pressure. The external id fills in automatically.

Leave the unit, the value type and any time range out of the name, they all have their own fields.

Pick a unit

Click + to choose: bar, °C, kWh. The unit is what makes a number readable and comparable across series, and picking one lets DataHub suggest the most storage-efficient value type for it.

Pick a value type

How each reading is stored. See the table below, this is the one that cannot be changed later.

Say whether it is a step series

A step series holds its value until the next reading (a set-point, a mode, a state). A non-step series is interpolated between readings (a continuous physical measurement). This affects how charts and aggregates read the gaps between samples.

Connect it to a resource

Click + on the relation field, pick the resource this series measures and a relationship. Do not skip this. An unconnected series is a column of numbers; a connected one is a property of a real thing. The linking step is contextualization, and it is what makes the series part of the model rather than beside it.

Assign a data set

Which container it belongs to, and therefore who can read and write it.

Value types

The value type sets how each datapoint is stored, its size, precision and compression. It is fixed when the series is created and cannot be changed later, so pick the most compact type that still represents your data honestly.

Value type	Use it for	Notes
Big Integer	Counters and tallies, whole numbers, no decimals	Most compact for integer values
Float	Analog measurements	64-bit floating point. Fast to aggregate, with tiny rounding
Float (32-bit)	Analog values where about 7 significant digits is enough	Half the storage of Float
Decimal (compact)	Small-range readings like temperature or percent	Exact to 4 decimal places, very compact
Numeric	Values that must be exact, billing, finance	Exact to 6 decimal places, large range
Text	States, labels, modes, non-numeric readings	Arbitrary text
Mixed (number or text)	Sensors that publish text statuses alongside numeric readings	Handles a `FAULT` string in a float series

Two practical notes:

Mixed exists for a real reason. Many industrial sensors publish FAULT, NODATA or similar strings in the same stream as their measurements. Without Mixed, those readings either break ingestion or get silently dropped, both worse than storing them.
Storage matters at scale. Choosing Float (32-bit) over Float across a million-series estate halves the storage bill for that data with no practical loss for most analog measurements. Choosing Numeric where Float would do costs several times as much.

The console shows this guidance behind the help icon next to the value-type field, so nobody has to remember it.

How often should you sample?

A question every new series raises, and one with a real answer from signal theory rather than a matter of taste.

The rule is the Nyquist–Shannon sampling theorem: to capture a signal faithfully, you must sample at more than twice the highest frequency you care about. Sample slower than that and fast behaviour does not just blur, it disguises itself as slow behaviour, an effect called aliasing. It is the same illusion that makes wheels appear to spin backwards on film: the camera samples too slowly for the rotation, so the motion reads as something it is not. An aliased sensor reading looks perfectly plausible on a chart, which is exactly what makes it dangerous.

Both panels sample the same fast signal, and every dot in both is a genuine measurement. On the left the rate clears the Nyquist limit and the shape survives. On the right the dots land exactly on a smooth slow wave that does not exist in the plant, and nothing downstream can detect the substitution, because the data is internally consistent.

Putting it to work:

Decide the fastest behaviour you actually care about

Not the fastest the instrument can measure, the fastest dynamics that matter to the question. A tank level drifts over minutes; a pressure surge lives in seconds; pump vibration lives in milliseconds.

Sample at least twice that fast, in practice more

Twice is the theoretical floor. Engineers normally use five to ten times the frequency of interest, so the shape of the behaviour survives, not just its existence.

Do not solve it by sampling everything at maximum rate

Storage grows with datapoint rate, and today measurement history is kept indefinitely, so an over-sampled series costs money forever. Why rate dominates the bill →

The practical rule of thumb per signal: fast enough that the behaviour you care about cannot hide, and no faster. A slow drift sampled every second is money wasted; a surge sampled every minute is a surge you will never see, or worse, will misread as a trend.

Finding series

The time series section has a dedicated search, and a filter panel to narrow by name, external id, unit, data set or the resource a series is attached to. Above a few thousand series, filtering by the resource is usually the fastest route, you know which equipment you care about, so start there and see what it emits.

To browse by data set instead, use the Time series tab of the data set section: click a data set and you get every series in it, including those of its sub-data sets. Seeing the time series in a data set →

Getting data in

A series is empty until data flows in. Datapoints arrive three ways:

From your systems, through the API. The normal production path. Your integration team writes to the API, or uses the SDK for Java, Python or Rust. Developer documentation →
From a change-data-capture integration. For source systems whose database can be tailed directly. Change data capture →
By hand, for testing. Copy your access token from the user menu at the top right and post a few readings.

Whichever route, the token is a credential, treat it like a password, and prefer a dedicated service account for production ingestion rather than a person's token. Users and access →

Editing a series

Name, unit, metadata, relationships and data set can all be changed after creation. The value type cannot: it defines the physical storage of every datapoint already written. The external id should be treated the same way, although the platform does not enforce it: integrations match on it, and a change breaks them silently. Why →

If a value type turns out to be wrong, the fix is to create a new series with the right type and migrate. That is deliberately a little painful, which is the point of choosing carefully at creation.