A dbt mart and a semantic model differ in that…

A dbt mart is a table; a semantic model declares which columns are measures, entities and joinable dimensions and is queried via the semantic-layer engine

They are the same thing under different vendor branding

A dbt mart is a table; a semantic model declares which columns are measures, entities and joinable dimensions and is queried via the semantic-layer engine

Marts are SQL; semantic models are NoSQL

Marts live in the cloud; semantic models live on-prem

Where the Semantic Layer Sits in the Modern Stack — Semantic Web Academy

Overview

A layered architecture, with hard boundaries

The healthy modern data stack has five layers, each with one job:

Layer	Job	Examples
Sources	Capture events / records	Postgres, Salesforce, Stripe, Segment
Warehouse	Store at scale	Snowflake, BigQuery, Databricks, Redshift
Transformation	Raw → staging → marts	dbt, SQLMesh, Dataform
Semantic Layer	Metrics, dimensions, joins as code	dbt SL, Cube, LookML, AtScale, Stardog
Consumers	Surface answers	Looker, Hex, Streamlit, embedded React, LLM agents, reverse-ETL

The no-overlap rules

Marts ≠ semantic models. A dbt mart is a table; a semantic model declares which columns are measures, which are entities, and which dimensions are joinable. The semantic layer reads marts; it does not replace them.
Semantic layer ≠ dashboard. The dashboard renders answers; the semantic layer produces them. A dashboard that embeds its own SQL is a dashboard bypassing the semantic layer.
Semantic layer ≠ catalog. A data catalog (DataHub, Atlan, OpenMetadata) describes what exists; the semantic layer says what every term means and computes to. They complement each other.

Most semantic-layer projects fail by collapsing two of these layers. Keep the boundaries crisp.

Wholesale market → station → recipe book → plate

Think of the warehouse as a wholesale food market, marts as prepared ingredients on a chef's station, the semantic layer as the kitchen's printed recipe book, and dashboards as plated dishes at the pass. A bistro that lets every cook improvise from the wholesale crates ends up with five different versions of the same risotto. A bistro that prints one recipe book and bans cooking off-book ships consistent dishes — and trains new cooks faster.

Reflect

Look at your own stack today. Pick any consumer surface (a dashboard, a notebook, an embedded chart). Trace its SQL backwards. Does it stop at the semantic layer or does it reach all the way down into a mart or a raw table? Every reach-around is a future five-revenues problem.

›How many of your dashboards bypass the semantic layer today (reach into marts directly)?
›What is the smallest set of dashboards you could re-point at the semantic layer this quarter to prove the model?

Reading in progress · 0 of 1 activity done