Data freshness is the age of the observations a metric was computed from. It is a property of the inputs, not of the metric, and it is routinely confused with when the metric was last calculated.
A metric recomputed an hour ago from a source that last reported in April is one hour old and four months stale. Only one of those numbers tells you anything useful.
Because it presents as stability.
When a source stops reporting, the metric does not error. It carries the last known value forward and draws a flat line. A flat line reads as a well-controlled process. Nobody investigates a metric that has stopped moving in the right direction.
Meanwhile the underlying condition can be deteriorating for months with nothing on the chart to indicate it.
Compare this to a source that starts returning garbage. That gets caught in days, because the number looks wrong. Silence is worse than noise.
Observation age per source, tracked as its own series alongside the metrics that depend on it.
A staleness threshold per source, derived from its expected reporting frequency. A daily source silent for three days is a problem. A quarterly source silent for three days is not.
Contribution volume per source per period. A source still reporting but returning a third of its usual records is failing partially, which is harder to spot than failing completely.
Freshness belongs next to the value, not buried in a data quality report nobody opens.
Where a value is being carried forward from an old observation, that should be visible to the person reading it. A quarterly figure from an audit closed in April is perfectly legitimate to show in July. Showing it without its date is not.
Metric Maestro records the observation timestamp for every fact it stores, including manually entered ones, and tracks ingestion runs per source.
That means the age of the data behind any value is available rather than inferred, and a source that has quietly stopped contributing shows up as an ageing input rather than as a stable-looking chart.
From the blog