Medallion architecture explained
Medallion architecture organizes a lakehouse into three layers, usually called bronze, silver and gold. Each layer makes data a little cleaner, better structured and more trustworthy than the one before.
The three layers
| Layer | What it holds | Who uses it |
|---|---|---|
| Bronze (raw) | Data exactly as it arrived from the source, plus load metadata. Append-only, kept for history and replay. | Data engineers |
| Silver (cleaned) | Deduplicated, validated, standardized data with consistent types and keys, joined and conformed across sources. | Engineers, analysts, data scientists |
| Gold (business-ready) | Aggregated, modeled data shaped for a business purpose: metrics, dimensional models, features, data products. | Analysts, dashboards, applications, AI |
Why layer at all?
- Replay and recovery. Because bronze keeps the raw data, you can rebuild silver and gold when logic changes or a bug is found.
- Clear responsibility. Each layer has one job, so problems are easier to locate.
- Progressive quality. Checks get stricter as data moves up, and consumers know what level of trust each layer offers.
- Reuse. Many gold datasets can be built from the same well-cleaned silver tables.
A simple example
Imagine order data arriving from an e-commerce system:
- Bronze: every order event lands as received, including duplicates and malformed records, with the time it was loaded and the source it came from.
- Silver: duplicates are removed, dates and currencies are standardized, invalid records are quarantined, and orders are linked to clean customer and product records.
- Gold: daily revenue by region, customer lifetime value, and a product-performance table that the dashboard and a forecasting model both read.
Governance and lineage across the layers
The layers are also a natural place to build trust:
- Lineage: record which bronze sources and which transformations produced every silver and gold table, so any number can be traced back.
- Access: raw bronze often contains sensitive fields; restrict it tightly and expose masked or aggregated data higher up.
- Quality: attach checks at each boundary (schema at bronze, rules at silver, reconciliations at gold).
- Ownership: make it clear who owns each gold dataset and the definition behind it.
Common mistakes
- Treating it as three copies of everything. Move data up only when it adds value.
- Business logic in bronze. Keep bronze raw; transformations belong in silver and gold.
- Gold built straight from bronze. Skipping silver duplicates cleaning logic across many tables.
- No owners for gold. A business metric without an owner slowly drifts out of meaning.
- Making three layers a rule. Some domains need more stages, some fewer. The principle is progressive refinement, not a fixed number.