The Medallion Architecture: Bronze, Silver, Gold Explained

What is the medallion architecture?

Popularised by Databricks, the medallion architecture is a data design pattern that structurally organises data in a lakehouse across three layers of progressively refined quality. Also known as the multi-hop architecture, it has become the de facto standard for building production-grade lakehouse platforms on Databricks, Microsoft Fabric and other modern data platforms.

The core principle is simple: data flows through a series of transformations, each improving its quality, structure and business value. Each layer serves a distinct purpose, enforces clear data contracts and provides a natural checkpoint for debugging, auditing and governance.

Bronze — the raw layer

Bronze holds the raw, unprocessed data ingested from source systems. It's append-only and immutable — a historical record of exactly what the source system sent, when it was sent and how it arrived. Think of it as your landing zone with schema-on-read.

The bronze layer serves several critical purposes. It provides a reliable recovery point if downstream transformations fail or produce unexpected results. It creates an auditable trail of source data for compliance and regulatory purposes. And it decouples source systems from the lakehouse — sources can send data at their own pace without worrying about downstream processing.

Silver — the cleansed layer

Silver is where data is filtered, cleaned and conformed. Joins happen here, duplicate records are deduplicated, data types are enforced and business logic is applied to transform raw data into a consistent, queryable form. This is your conformed dimensional model.

Silver tables represent the single source of truth for each business entity — customers, products, transactions. They resolve conflicts between different source systems, apply master data management rules and enforce data quality standards. Most analytics and data science work happens against the silver layer.

Gold — the curated layer

Gold is the presentation layer — aggregated, business-ready data shaped for specific use cases: executive dashboards, ML feature stores, downstream APIs and regulatory reports. Data here is highly refined and often denormalised for performance.

Gold layer tables are project-specific and consumer-oriented. A sales dashboard might have a gold table that pre-aggregates revenue by region, product category and month. An ML pipeline might have a gold table with engineered features ready for model training. The gold layer is where business users and analysts spend most of their time.

Why it works

The separation enforces clear contracts between layers. When a pipeline breaks, you know exactly which layer failed and can trace the issue to a specific transformation. This isolation also enables parallel development — one team can work on silver transformations while another builds gold aggregations, as long as they respect the data contract at each boundary.

The medallion architecture also supports incremental processing. Instead of reprocessing entire datasets, you can upsert only new or changed records at each layer. Delta Lake's change data feed and MERGE operations make this pattern straightforward to implement and maintain.

Common pitfalls

• Skipping bronze and going straight to silver — you lose your audit trail and recovery point
• Over-processing in gold — gold should aggregate and shape, not re-clean data that should have been handled in silver
• No clear data contracts between layers — without explicit schemas and quality expectations, layers become coupled and brittle
• Ignoring incremental loading — full refreshes work for small datasets but become unsustainable as data grows

Getting started

Start with bronze and silver before gold. Define your data contracts early. Implement data quality checks at each layer boundary using Delta Live Tables expectations or Great Expectations. And document your lineage — when something breaks at 3am, you'll be grateful you did.