Review a Power BI Release Before It Ships
A report change is easy to review: open it and look. A semantic model change is not. A renamed measure, a partition pointed at a different server, or a relationship switched to filter both ways can all look like nothing in the report until someone’s number moves. This is the review I run before publishing a model, using the free tools on this site.
It assumes the model is saved as a Power BI project with the semantic model in TMDL, so there is a definition folder of plain-text files.
1. Compare It with What Is in Production
Open Model Compare and load two versions: the one in production (the last release tag, or a fresh export) and the one you are about to publish. You get every table, column, measure, partition, shared expression, calculation item, relationship and role that was added, removed, renamed, moved or changed, with a line diff of the DAX and M.
Read it in this order:
- Partitions and expressions. A changed
Sql.Databaseserver or a parameter set to a dev value is the most expensive mistake to ship, and the easiest to miss. - Relationships and roles. A cross-filter change or an edited role filter changes numbers or access without touching a single measure.
- Measures. Read each DAX diff. A rename shows as one removed and one added (a change of case alone shows as renamed). Reports in other files and Excel workbooks that use the old name break, so check the removed entries against everything built on the model.
The page has no network access, so a production model never leaves your machine.
2. Read the Checks
Drop the new version into the Model Documenter. Its Model Checks list runs every DAX rule on each measure, calculated column, calculation item and security filter, and the model rules on relationships, columns, sources and roles.
Treat warnings as “explain or fix before release” and notes as “worth a look”. The ones I would never ship without a reason are a row filter cleared inside an iterator, a relationship that filters both ways, a role with no filters, and a source that points at a local file or localhost.
3. Keep a Markdown Dictionary in Git
Download the Markdown version of the dictionary and commit it next to the model. It lists every table, column and measure with its DAX, what each measure uses and what uses it, and the checks. The next time someone opens a pull request, the diff of that one file shows what changed in words a reviewer can read, alongside the TMDL diff that shows exactly how.
What This Does Not Cover
None of these tools run the model. They do not refresh data, evaluate a measure or render a visual, so they cannot tell you that a number is right, only what changed and what looks risky. Before publishing, still open the reports on the new version and check a handful of numbers against the old one: one card per page is usually enough to catch a broken relationship.