Problem
Prada Group had thousands of retail KPIs spread across Excel files and database systems, each with its own definition and calculation function. Teams could not reuse logic or agree on what a number meant. The KPIs also had to reach very different consumers: BI dashboards, data mesh teams, analysts, agents and apps. One of them, a clienteling app used by sales advisors in stores, needs the numbers while a customer is standing at the counter. Serving KPIs through Databricks SQL warehouses took 1 to 2 seconds end to end, too slow for that app.
Example from the article: a client advisor opens the clienteling app and needs the store's monthly sales quantity at once.
Solution in brief
Define every KPI once in a YAML descriptor, governed in Unity Catalog and tied to a business taxonomy. Compute KPIs in Lakehouse jobs from unified entities, sync the results to Lakebase, and expose them through taxonomy-oriented functions behind a Data API. One call format serves every channel, and the underlying tables stay hidden.
Questions this case answers
- How to define each business metric once in a governed layer and serve it to many consumers through one call format with sub-second latency
- How to organize thousands of metric definitions that live in scattered files and databases into one shared taxonomy
- How to define KPIs once in a governed semantic layer and reuse them across BI, apps and agents
- How to serve precomputed KPIs to an in-store app in under 200 ms
Tools and methods
- Unity Catalog — governs the taxonomy, KPI definitions and mapping tables
- YAML KPI descriptor — formula, dimensions, filter and time period for each KPI
- Lakehouse jobs — compute KPIs from silver-layer entities into KPI store tables
- Lakebase (Postgres) — cached KPI stores, PL/SQL functions and the Data APIs layer
- Change data feed — incremental sync of KPI stores into Lakebase
Google or a one-time email code. No password.