Project summary · Microsoft Fabric · PySpark · T-SQL · Power BI
End-to-End Retail Analytics on Microsoft Fabric
A full medallion lakehouse in Fabric — raw landing, PySpark cleanup, a T-SQL star schema in the Warehouse, and a deployment pipeline that promotes it all between workspaces.
Why I built it
Job descriptions kept asking for Fabric plus a real promotion path between environments — not a single workspace with everything dumped in it. This project builds the whole stack the way a team would run it, including source control and CI/CD.
Questions it answers
- Where does raw retail data land, and what makes a layer safe to query?
- How do you move a validated model from a development workspace into production without rebuilding it by hand?
- What does the reporting layer look like when the star schema is enforced in the warehouse rather than in Power BI?
How I built it
- Bronze landing. Raw retail source files land in a BronzeLH lakehouse, untouched, so any downstream mistake can be replayed from source.
- Silver cleanup in PySpark. A notebook standardises types, handles nulls and de-duplicates into SilverLH — the layer everything else reads from.
- Gold star schema in T-SQL. A fact table plus three conformed dimensions built in the Fabric Warehouse, with an aggregated view on top for the report to hit directly.
- CI/CD. A Fabric Deployment Pipeline promotes the workspace from GlcAnalyticsWS to CICDWS, so environment promotion is a button rather than a rebuild.
- Source control. The workspace is Git-backed, so the pipeline definitions and notebooks live in the repo alongside the documentation.
How I know it works
- Each layer is queryable independently, so a bad number can be traced back one hop at a time.
- Gold is a declared star schema — the fact grain is fixed in the warehouse, not implied by a Power BI relationship.
- The deployment pipeline was run end to end between two real workspaces, not just configured.
Honest notes
The dataset and the overall pattern come from Ansh Lamba's Fabric tutorial, which is attributed in the repo README. The pipeline build, the cleanup logic, the data model and the CI/CD setup are my own implementation.