The semantic layer
Two dashboards disagreeing about revenue is not a BI problem. It is a definition problem, and it happens because the definition lives in whichever tool computed it.
The semantic layer puts the definition in the catalog, once, and lets every tool ask for it.
What a definition carries
| Element | What it is |
|---|---|
| Measures | The quantities: revenue, active users, remaining useful life |
| Dimensions | What they are sliced by |
| Filters | The qualifications that are part of the definition rather than added per query |
| Hierarchies | Resolved into drill paths, including ragged and parent-child shapes |
| Metric views | A metric exposed as something queryable |
| Benchmark definitions | What a measure is compared against |
| Glossary links | A measure bound to the business term it implements |
A semantic cache sits over it, so a repeated metric query does not re-derive work that has not changed.
It generates native BI configuration
This is the part that decides whether a semantic layer gets adopted or ignored.
Rather than asking analysts to abandon their tools and query through a new surface, the platform emits native configuration for the tools they already run: Tableau, Power BI tabular models, Superset and Grafana.
So the definition is authored once in the catalog and materialises into each tool's own vocabulary. The analyst keeps their tool. The definition stops being per-tool.
Alongside generation, the platform also imports semantic models. A canonical intermediate representation parses and serialises TMDL, BIM, PBIX, LookML, Cube, MetricFlow, Mondrian, MDX, transformation projects and Power Query, with DAX and M toolchains carrying linting, best-practice analysis and model-health scoring.
The practical consequence is that an existing estate's semantic models can be read in, assessed, and migrated rather than rebuilt from a blank page.
Governance comes along
A metric query is a query. It compiles through the same plan as any other read, so it carries row filters, column masks, a classification on the result, a lineage record and a cost.
That means a measure is not a way around policy. Two analysts with different clearances asking for the same metric get different, correctly filtered answers, rather than one of them getting a number they should not have because it arrived pre-aggregated.
Dimensional modelling underneath
Where a star schema is the right shape, the platform models it explicitly rather than leaving it implicit: slowly changing dimensions in types 1, 2 and 3, conformed dimensions shared across subject areas, grain enforcement so a fact is at the grain it claims, ragged and parent-child hierarchies, materialisation, an aggregate navigator that routes a query to the most appropriate aggregate level, and incremental refresh.
Industry packs ship complete dimensional models built on exactly this, which is the fastest way to see what the layer is for.
Related
- Data modeling, transformation projects that produce the tables underneath.
- Reporting and BI, every surface a reporting tool connects over.
- The catalog and lineage, where definitions live.
- Search and discovery, finding a measure rather than re-inventing it.