Data
Build warehouse models
Staged, tested and documented models over raw exports.
Fill it in
What lands in the warehouse.
Uniqueness, not-null and the relationships.
What each one means, not what it selects.
Your prompt
Build models over . Stage first: one staging model per source, renaming and typing only. Business logic in a model that also cleans is the pattern that makes a warehouse unmaintainable, because nobody can tell which layer a wrong number came from. Test the things that break silently: uniqueness on the key, not-null on what joins, accepted values on enums, and the relationship between the fact and its dimensions. An untested model is a view somebody trusts. Document what each model MEANS. "Active customers" is a definition, and if it is not written down there will be three of them within a year. Name things so a business reader can read the SQL.
Continue and pick your folderOpens with everything above already filled in.
Why this works
Business logic inside a model that also cleans is what makes a warehouse unmaintainable, because nobody can tell which layer produced a wrong number. This stages first, tests the things that break silently, and writes down what "active customer" means before there are three definitions.