Without a shared convention, every team that touches a lakehouse ends up inventing its own folder structure and its own definition of "clean" — one analyst's cleaned table is another's half-finished one. The Medallion architecture (popularized by Databricks, but not tied to any one vendor) fixes this by giving every table a clear place in a three-stage progression, so anyone opening the platform for the first time already knows what to expect from a table based on which layer it's in.

Bronze tables are the source data landed as-is, with minimal transformation — usually just adding ingestion metadata (load timestamp, source file name). Nothing is deduplicated or reshaped here. If a downstream bug corrupts a silver table, bronze is what you replay from.

Silver applies the transformations everyone actually needs: deduplication, type casting, joining reference data, filtering out known-bad records. Silver tables still sit close to the source grain — one row per order line, not per customer per month — but they're now trustworthy enough that any team can build on them without redoing the cleanup themselves.

Gold tables are aggregated and shaped around a specific business question: daily revenue by region, active users by cohort. They're intentionally denormalized and disposable — a gold table can always be rebuilt from silver, so it's safe to optimize purely for the dashboard or model that consumes it.

# Bronze -> Silver: dedupe and enforce types, nothing else
silver_df = (
    bronze_df
    .dropDuplicates(["order_id"])
    .withColumn("order_date", F.to_date("order_date"))
    .filter(F.col("order_id").isNotNull())
)
silver_df.write.format("delta").mode("overwrite").save("s3://lakehouse/silver/orders/")

Nothing here is business-specific — it's the same handful of checks every team would otherwise write themselves, done once, in one place everyone can rely on.