Golden Data

Issue #16 · August 26, 2026

Snowflake And Databricks Now Ingest Your Power BI Model. Most Of Your Skill Is In The Part They Skip.

Paul BrownWrites Golden Data from Ogden, Utah7 min read

On August 18 Snowflake made a Power BI file an input. You upload a .pbix to Semantic View Autopilot. It reads the model, matches it to tables already in your account, and writes a semantic view. Relationships carry over. Column descriptions carry over. Metric definitions carry over. Tableau workbooks have been making the same trip since at least January. The larger installed base joined this month. Databricks has been doing the same through Genie Code since June.

Databricks, June. Genie Code takes a Tableau or Power BI file and writes the measures and dimensions into Unity Catalog metric views. The docs say to attach a screenshot of the original dashboard so the agent can check its numbers against the picture. Two warehouses, one summer, the same upload button.

For most of your career the semantic model was the thing you could not take with you. The reports could be rebuilt anywhere. The definitions lived inside the tool, in the tool’s language, and every migration stalled at the point where someone had to retype them. That stall was your job security whether you named it or not. This month one vendor turned it into an upload button.

Then read the support table, because Snowflake published one, and the honesty is in the exclusions. Report-level measures do not fully port. Time intelligence does not port: PREVIOUSMONTH, SAMEPERIODLASTYEAR, TOTALYTD. On the Tableau side the same documentation lists Level of Detail calculations as unsupported. Two tools, two carve-outs, one pattern. What survives the trip is the part that was always describable in plain SQL. What stays behind is the part written in the tool’s own dialect.

That is a sentence about your résumé, not about Snowflake.

Four things moved this month, and they moved in the same direction.

Snowflake, August 18. The warehouse reads Power BI’s files. Power BI cannot read the warehouse’s. In Snowflake’s own hands-on lab, the path back to Tableau is a custom UDF someone wrote to emit a .tds file. One direction is a product. The other is a workaround.

Microsoft, August 18, in the monthly Power BI update. By the end of August, Fabric App consumers need only Read permission on a semantic model instead of Build. Read got cheaper. Build, the permission you need to take a model out of the building, is where it was.

Apache Ossie, incubating. The Open Semantic Interchange spec moved to the Apache Software Foundation in June with four merged converters: dbt’s MetricFlow, GoodData, Salesforce, Apache Polaris. Every one an engineering-side tool. No Power BI converter. The neutral standard reaches the warehouse crowd first and the largest installed base of semantic models last.

The benchmark that does not exist. Two studies this spring agree that a semantic layer is worth about twenty points of accuracy to a model answering questions over data. Rumiantsau and Fokeev at Cube put 99 retail questions to Claude Opus 4.7, Claude Sonnet 4.6, and GPT-5.4, twice each: once with the warehouse schema, once with the schema plus a 4 KB markdown file an analyst wrote by hand. The file was worth 17 to 23 points. With it, the three models are indistinguishable. Without it, they are also indistinguishable, twenty points lower. The model did not matter. The page did. dbt Labs found the same shape on its own blog three weeks earlier, from a higher floor. Both are vendor studies. Neither measures how much of a semantic model survives being moved. It is the number a practitioner needs most.

Here is what the pattern means for a senior analyst or a BI lead. The half of your model that ports is the half a warehouse engineer could have written. The half that does not is the half you learned from the tool’s documentation, the certification, the forum thread with the clever DAX trick. Read that either way. It is a moat, or it is a skill priced by one vendor. A skill that cannot travel is a skill with one buyer, and one buyer sets the price. Open source is the bridge.

The action. One, and it fits in a Friday afternoon.

Export the measure list from the model your forecast depends on. In Power BI that is DAX Studio or Tabular Editor. In Tableau it is the calculated fields in the workbook XML. Tag every measure that calls a time intelligence function or an LOD expression. Count them. Then take the ten the Monday call cannot run without and rewrite each one where the warehouse can read it: a dbt metric, a semantic view metric, a dated CTE in a view with an owner and a grain. Do not migrate. Nobody asked you to migrate. Write the second copy and check whether it reproduces last quarter’s reported number.

What you learn is not whether the tool is dead. It is which of your definitions you can explain without the tool open. The ones you cannot are the ones you are renting.

What stayed behind was your dialect. What goes forward is your judgment. The next four issues take that apart, one skill at a time: attention, decisions, vision, systems. This one is the door.

The next four

Attention. The Warehouse Can Read All 300 Of Your Measures. It Still Can’t Tell Which One The Room Decides On. The arXiv study’s semantic file was 4 KB. Not 400. Someone chose what went in it, and the choosing was worth twenty points of accuracy. Ingestion ports every definition with equal weight, which is exactly how nobody in a forecast call thinks. Salience does not port. It never lived in the model; it lived in whoever knew which number the CFO looks at first. The action: write your own 4 KB file. Twenty metrics, in prose, each with the decision it feeds and the person who makes it. A metric with no decision goes in a second file called “computed.”

Decisions. Two Surfaces Now Answer Every Question Fluently. Somebody Still Has To Say Which One Is Right. A parity check surfaces disagreements; it does not settle them. When Power BI says one net new ARR and the semantic view says another, the meeting does not need a better query. It needs a person who will put their name on a definition, say it out loud, and take the call when finance disagrees. That person used to be whoever owned the tool. Now it is whoever owns the sentence. The action: take the ten metrics from this issue and fill in the owner column with a name. If the name is not yours and the owner does not know they own it, you have found the real gap.

Vision. Ingestion Moved The Formula. It Left The Question The Formula Was Built To Answer. Every measure in your model exists because someone asked something in 2021. The warehouse received the arithmetic and none of the intent. And the model’s next reader is not a human on a dashboard. It is an agent that will be asked questions nobody has asked yet, in words nobody chose. Modeling for a reader that does not exist yet is the skill, and it is the one the certification never covered. The action: for each of the ten, write the question a VP will ask in four quarters that this metric cannot answer. That list is your next model.

Systems. The Model Lives In Two Places Now. The Job Is Keeping Them From Drifting Apart. A contract per metric. Grain, filter logic, owner, a hash, a last-verified date. Not a dashboard, a table with a tolerance. The skill is thinking in pipelines and drift instead of visuals and refresh. The action: for the ten, build a one-page table that shows both surfaces’ values side by side with a delta. Run it by hand on Monday morning for four weeks before you automate anything. The weeks it is green are the weeks you learn nothing. Watch for the first red.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown