medallion architecture is databricks' name for bronze-silver-gold. raw, clean, business-ready. data flows one way and gets better at each step in the lakehouse.
bronze is where you land everything as-is. logs, apis, cdc dumps. schema-on-read. cheap, messy, complete.
silver is where you validate, dedupe, type, join. still normalized. this is your enterprise data layer.
gold is where you denormalize, aggregate, build star schemas and feature stores. this is what the business actually queries.
inmon and kimball are older answers to a different question: how do you model data for analytics?
inmon goes top-down. build a normalized enterprise warehouse in third normal form. strict governance. single source of truth. then spin off subject-specific data marts. heavy etl, slow start, clean result.
kimball goes bottom-up. build dimensional data marts straight from sources. star schemas aligned to business processes. fast time-to-insight, less central control, easier for ad-hoc analysis.
here's how they fit inside medallion:
inmon's normalized warehouse maps cleanly to the silver layer. his data marts become gold. the lakehouse gives inmon something he never had: cheap raw storage in bronze, and schema evolution without migration hell.
kimball's star schemas live naturally in gold, with bronze-to-silver acting as his staging area. his bus architecture of incremental dimensional loads maps well to incremental etl into silver and gold.
medallion answers "where does data live and how does it flow." inmon and kimball answer "how do i model it at each stage." silver can be 3nf. gold can be star schema. or both. the lakehouse doesn't force the choice early.
defer schema decisions. land raw cheaply. model when you actually know what you're building.
the thought is mine. the words are written by janis, my hermes agent.