Semantic layer and metrics store

Problem class

Executives receive diverging answers to simple questions ("What was revenue last month?") because Finance and Data Science define "revenue" differently in their respective tools. Tableau dashboards, Power BI reports, dbt models, and ad-hoc SQL views each compute the same metric with subtly different logic. Every BI tool proliferates its own calculated fields. When definitions change, they must be updated in every tool individually. NL-to-SQL systems fail because there is no single source of truth for metric semantics.

Mechanism

A semantic layer sits between the physical data store and all consuming tools (BI dashboards, experimentation platforms, ML feature pipelines, NL-to-SQL interfaces). Metric definitions — including grain, filters, time logic, and business rules — are declared once in the semantic layer and served to all downstream consumers via a unified API (SQL dialect, GraphQL, or REST). dbt Semantic Layer, Cube, Looker/LookML, and AtScale all implement variants of this pattern. Consumers query "Revenue" rather than writing SQL logic; the semantic layer compiles the query to the physical warehouse. Caching and push-down optimization prevent per-query recomputation.

Required inputs

• Governed ELT models as the physical data foundation
• Metric glossary co-created with Finance, Product, and Marketing stakeholders
• Semantic layer tooling (dbt Semantic Layer, Cube, Looker, AtScale)
• BI tool connectors consuming from the semantic layer API
• Change management process for metric definition governance

Produced outputs

• Single, auditable metric registry accessible to all BI tools
• Consistent metric calculations across dashboards, experiments, and ML features
• Dramatically reduced data mart proliferation (one customer reduced data marts by 93%, saving $10M annually)
• Improved NL-to-SQL accuracy (from ~10% to ~90% with semantic context)
• Reduced time to answer contested business questions

Industries where this is standard

• Data-forward technology companies (Airbnb, Uber, Spotify) requiring metric standardization across product, finance, and engineering
• E-commerce/retail/CPG with consistent revenue, margin, and customer metrics across regions
• Financial services with regulatory reporting requiring auditable, consistent metrics
• Media and publishing (Forrester study customer: 25% CPM reduction, 30% revenue improvement)
• Telecommunications with certified datasets as single source of truth

Counterexamples

• Teams with fewer than 5 contested metrics: The governance overhead isn't justified until metric disagreements are actively causing business problems.
• Schema-driven rather than business-driven design: Building semantic models based on database schema rather than a metric glossary co-created with stakeholders produces technically correct but business-irrelevant models.
• Building the complete semantic layer before launching: Start with 3–5 core contested metrics (Revenue, Active Users, Churn) rather than attempting enterprise-wide coverage from day one.

Representative implementations

• Airbnb (Minerva) manages 12,000+ metrics and 4,000+ dimensions with 200+ data producers. Before Minerva, executives received diverging answers to simple questions from Data Science and Finance teams. During COVID-19, the platform's prototyping tool "reduced iteration from days to minutes."
• Bilt Rewards achieved an 80% reduction in analytics costs by migrating from embedded BI to the dbt Semantic Layer's GraphQL endpoint, with $20,000/month in BigQuery savings and 99% reduction in data volume scanned. A 3-person analytics team now serves 10,000+ external B2B data consumers.
• Looker/LookML (Forrester TEI) delivered 486% ROI over 3 years ($8.9M NPV), a 99%+ reduction in reliance on technical teams for analytics, and a 100% increase in completed A/B tests (250 → 500 annually). One customer reduced data marts by 93%, saving $10M annually.
• Uber (uMetric) found that standardizing impression definitions reduced transient-view impressions by up to 30%, providing far more accurate measurement.

Common tooling categories

Semantic layer engine (dbt Semantic Layer / Cube / Looker LookML / AtScale / Metriql) + BI connector (Tableau / Power BI / Metabase / Superset) + metric glossary management + caching layer + API gateway for downstream consumers.