Semantic layers are finally getting opinionated enough to be useful
Summary
Semantic layers are finally getting opinionated enough to be useful
The semantic layer has been around forever. It's the little translation booth between raw data and the decisions people make with it. For years it sat there like the seatbelt in a self-driving car demo, technically present, mostly ignored. AI is what finally made it necessary.
The numbers are grim. 40% of Databricks users still won't touch dbt. Every BI tool in the org lugs around its own definition of "revenue." You wind up with dozens of dashboards that quietly disagree, like a room full of people who all swear they said the same thing.
AtScale, Stardog, Databricks Unity Catalog Metrics, and the rest let you define a metric once and reuse it everywhere: SQL, DAX, MDX, Python, AI agents. Everyone sells it as no-code BI. The thing it actually fixes is drift. Your human analysts, your ML pipeline, and whatever LLM you point at the warehouse all read one "revenue" the same way, so dashboards and model training data agree for once.
The AtScale and Databricks "Semantic Lakehouse" nails this. Nothing gets shuffled around, aggregates come for free, one set of metric definitions, wired right into Unity Catalog and Spark. AI finally gets solid ground to stand on.
My new TIL post, "Semantic Layer Solutions in Modern Data Architecture" covers vendors (AtScale, Stardog, Timbr), integrations (Unity Catalog Metrics, Power BI), and the Databricks + AtScale partnership that puts real weight behind the phrase "semantic lakehouse." It walks through what a semantic layer actually is, why 40% of Databricks users still don't touch dbt, and how AtScale and Databricks Unity Catalog Metrics chip away at the "truth problem" in analytics.
This goes way past BI, though. Get "revenue" to land on one number no matter how you ask for it, and you've poured a foundation that people and machines can both stand on. The dashboards inherit that consistency for free.
"AtScale's main selling point is that it stops data from moving..." It queries data in place within Databricks, creates and manages aggregates independently, accelerates performance through intelligent caching, and maintains a single source of truth without duplication. (AtScale x Databricks blog)
Pull data into a BI cube and you've started the drift. AtScale's Databricks integration closes that loop, wiring technical lineage (through Unity Catalog) to the business-facing meaning of the words. Foundation work. AI architectures get poured on top of it.
Then you point the same layer at the model you're training and testing. The "revenue" that feeds your dashboards should feed your model features and your eval metrics too. Otherwise the AI learns the business from one definition while your executives run on another, and nobody notices until the numbers fight in a meeting.
I'd push back on a few things first. "Single source of truth" is just a slogan until you hard-gate BI against the layer. Let people keep aiming their own tools straight at warehouse tables and you've got two truths sitting in the same building. The pitch says MDX, DAX, and LookML all say the same thing. On real workloads they don't quite. Edge-case functions and time-intelligence logic refuse to map one-to-one, so pad your schedule for the testing. Vendor lock-in costs real money, too. A small BI surface can probably ride Unity Catalog Metrics and dbt/MetricFlow-style semantics. Once you sprawl across surfaces and the org politics go decentralized, AtScale starts earning its keep.
The full TIL has the vendor details, query protocols, and implementation patterns.