The medallion architecture organizes a lakehouse into three quality tiers - bronze, silver, and gold - through which data flows as it's progressively cleaned and refined. Popularized by Databricks, it gives an otherwise sprawling data lake a predictable structure: raw data lands in bronze, gets conformed in silver, and is shaped into business-ready tables in gold. The value isn't the medal names; it's the discipline of separating ingestion, cleaning, and serving into distinct, reprocessable layers.

The three layers

Bronze holds raw data exactly as ingested - append-only, schema-on-read, with source metadata and ingestion timestamps. It's the system of record you can always replay from. Silver is cleaned, deduplicated, type-cast, and conformed: bad records filtered, keys standardized, late data merged, tables joined into a coherent enterprise view. Gold is business-level aggregates and dimensional models - the star schemas, KPIs, and feature tables that dashboards and ML read directly.

raw events ─▶ bronze (raw, append-only)
            ─▶ silver (cleaned, conformed, deduped)
            ─▶ gold   (aggregated, business-ready: facts, dims, KPIs)

Why layering pays off

Keeping bronze immutable means any bug in cleaning or modeling is recoverable - you reprocess silver and gold from bronze without re-ingesting from the source. Each layer has a clear contract, so a change to a gold aggregate doesn't require touching ingestion, and a new downstream use case can branch off silver without disturbing existing gold tables. Lineage runs cleanly in one direction, which makes debugging and auditing far simpler than a flat lake where raw and derived data are tangled together.

Practical cautions

The layers are a guideline, not a law: not every pipeline needs all three, and over-literal teams sometimes create needless copies or push transformations into the wrong tier. Heavy business logic belongs in the bronze-to-silver and silver-to-gold transforms, not smeared across all of them. The architecture pairs naturally with the lakehouse and with dimensional modeling in the gold layer - bronze and silver feed the facts and dimensions that gold exposes.

Frequently asked questions

What do bronze, silver, and gold mean? Bronze is raw ingested data, silver is cleaned and conformed data, and gold is business-ready aggregates and dimensional models. Data quality increases at each step.

Do I always need all three layers? No. Simple pipelines can collapse to two, and some add sub-layers. The principle - separate raw, refined, and serving - matters more than the exact count.

Why keep raw bronze data around? So you can reprocess everything downstream after a logic change or bug without re-pulling from source systems. Bronze is your replayable system of record.

Is medallion specific to Databricks? The term is Databricks', but the layered raw-clean-serve pattern is general and works on any lakehouse or warehouse stack.