A dimension is an attribute attached to a measurement that lets it be broken down. Business unit, region, environment, severity, asset class, owner. One metric plus four dimensions answers far more questions than four separate metrics.
Without dimensions, a coverage figure is one number for the enterprise. Someone asks which part of the organization is behind, and the answer requires a new metric.
With dimensions, the same figure can be cut by region, by business unit, by environment, by asset criticality, from one definition. The metric is defined once and sliced many ways, which is also what keeps the definitions consistent across those slices.
Validate values against a defined list rather than accepting whatever the source sends. This is the single most valuable control on a dimension.
Free-text values drift immediately. Three spellings of a department name produce three series that should have been one, and the split is invisible until someone notices the totals do not add up.
A lookup table that rejects unrecognised values pushes the problem back to the source, where it can be fixed once, rather than letting it corrupt every downstream breakdown.
Things people slice by, and only those. Attributes with a bounded set of values that someone will actually filter on.
Identifiers do not belong here. Asset ID and user ID have unbounded distinct values and belong to the underlying records, supporting drill-down rather than slicing. See cardinality.
Four or five per metric is usually the practical ceiling. Beyond that the combinations stop being examined, and the cost of computing and storing them stops being repaid.
If a metric seems to need eight dimensions, it is often two metrics that have been merged.
From the blog