Home › Data engineering › Metric views › 1 · The idea
Data engineering · BI and analytics · free guide · 1 · The idea
What a semantic model is, before any product
Forget every product for a moment. A semantic model is a written agreement about what business numbers mean, kept in one place, in a form a machine can turn into a query. When a sales manager asks for "net revenue by region last month", the semantic model knows which table, which column, which join, which rows to leave out and which rows that manager may see.
Kitchen, menu, diners. The menu names each dish and holds the recipe; diners never cook from the raw ingredients.
Why it exists
The problem it solves
Without one, every dashboard holds its own SQL. Over time the copies drift: one adds returns, one leaves out a dealer type, one joins a table that repeats rows. Then meetings start with "whose number is right?". With one, the recipe lives once, so the argument happens once, when the metric is defined, and never again in each meeting.
AI makes this sharper. A model writing SQL straight on raw tables guesses joins and filters, and guesses differently each time. A model that picks a named metric and named slices from a semantic model cannot invent a recipe. That is the main reason semantic models are back in fashion.
The parts
Ten parts every semantic model has
The ten parts. Every product on the next page builds these, in its own words.
| Part | What it is | Why it matters | Example (tyre sales) |
|---|---|---|---|
| 1 Grain | What one row of each table means. | Every number is computed from rows. If you don't know what a row is, you can't know what SUM means. | Invoice line, not invoice. |
| 2 Relationships | How tables join, and whether each join is one-to-one, many-to-one or one-to-many. | A wrong join repeats rows and inflates sums without any error. | Each line has one dealer; a dealer has many lines. |
| 3 Dimensions | The ways a number can be sliced, including time at several grains and hierarchies. | Business questions are 'X by Y': revenue by region by month. | region, dealer, segment, invoice month |
| 4 Measures | Aggregations over rows, without a fixed grouping. | The grouping comes from the question, so one measure serves every report. | SUM(net_amount) |
| 5 Metrics | Numbers built from measures: ratios, differences, last year, running totals. | Most arguments are about these, because they hide choices (ratio of sums or average of ratios?). | margin % = gross margin / net revenue |
| 6 Scope | Rules about what counts at all. | Cancelled orders, test dealers and internal sales must never sneak in. | status <> 'CANCELLED' |
| 7 Access | Who may see which rows and columns. | The same metric must give each person only their part. | North managers see North. |
| 8 Business names | Labels, descriptions, synonyms, formats, owner, certified or not. | People and AI find numbers by words, not column names. | 'zone' means region; show as ₹ with no decimals. |
| 9 Query interface | How a tool asks: metric + dimensions + filters in, an answer out. | Tools stop writing their own SQL, so they can't get the recipe wrong. | MEASURE(net_revenue) by region |
| 10 Speed layer | Caches and pre-computed aggregates, used only when they give the same answer. | Big tables and busy dashboards need answers in seconds. | a daily-by-region aggregate |
Four ideas to hold on to
What makes a semantic model good
- It is a contract, not a report. A report answers one question. A semantic model answers every question built from its parts, and promises that the same question always gets the same number.
- It holds meaning, not data. The data stays in tables. The model holds names, recipes and rules. Change a recipe once, and every tool sees the change.
- It sits on clean, atomic data. Feed it pre-summed tables and it can only slice what someone already summed. Feed it the finest grain any metric needs, and it can answer questions nobody planned.
- Its users are tools. Dashboards, spreadsheets, notebooks, Genie and agents ask it questions. That is why business names and synonyms are part of the model, not decoration.
Concept vs physical
Keep the idea separate from the product
The ten parts are the concept: the job and its logic. A metric view, a dbt semantic model or a Power BI dataset is the physical thing doing the job, like Java or Python doing the same algorithm. Learn the parts once and you can read any product: ask where it keeps the grain, how it declares joins, how it composes metrics, where access rules live and how tools ask it questions.
← Start here · Next: 2 · Three products →
FAQ
Questions people ask
What is the difference between a semantic model and a data model?
A data model shapes the tables: facts, dimensions, keys. A semantic model sits on top and adds meaning: named measures and metrics, allowed slices, scope rules, access rules and business names, plus a way for tools to ask for them.
Is a semantic layer the same as a semantic model?
In practice yes. 'Semantic layer' is the layer in the architecture; 'semantic model' is what you write in it. Some vendors also call it a metrics layer.
Why should gold tables be atomic under a semantic model?
Because a semantic model can only slice what is in its source. If gold is already summed by month and region, nobody can ask by dealer or by day, and ratios computed before summing come out wrong.
Sources
Checked against (11 Nov 2026)
- Databricks: Unity Catalog metric views and the YAML reference
- Materialization · Advanced techniques · Relationships · BI tools · Feature availability
- Row filters and column masks · ABAC evaluation through views (KB)
- dbt semantic models · MetricFlow open-sourced · Apache Ossie
Metric views change every few months. Confirm runtime minimums and preview status before a design depends on them.