Data Engineering › Storage, Formats & Lakehouse
Medallion Architecture (Bronze, Silver, Gold)
Layering data from raw to cleaned to business-ready.
Also known as: bronze silver gold, medallion layers, multi-hop architecture, data layers, bronze silver gold layers
Medallion architecture organizes data in layers of increasing quality: Bronze (raw), Silver (cleaned) and Gold (business-ready). The names and the pattern were popularized by Databricks, and the idea is widely used with any lake or lakehouse.
sources ─► BRONZE ─────► SILVER ─────► GOLD ─────► dashboards, ML, reports
raw, cleaned, aggregated,
as received deduplicated, modeled for a
typed, joined business purpose| Layer | Contains | Typical work |
|---|---|---|
| Bronze | Raw data as ingested, plus load metadata | Append-only; keep original (landing zone) |
| Silver | Validated, deduplicated, typed, standardized records | Cleaning, conforming keys, joining reference data (deduplication) |
| Gold | Curated tables for specific uses: facts and dimensions, aggregates, features | Business logic and metrics (dimensional modeling, data marts) |
Why layers help
- Reprocessing is easy. If Silver logic is wrong, rebuild it from Bronze.
- Quality improves in steps, and each step can be tested and owned.
- Different consumers use different layers: data engineers read Bronze, analysts mostly use Gold, and data scientists often use Silver.
- Clear contracts: Gold tables are what the business can rely on.
Practical guidance
- It’s a convention, not a rule. Some teams use other names (raw, staging, intermediate, marts, or two layers instead of three). Use whatever’s meaningful and consistent.
- Don’t over-layer. Every extra hop adds storage, runtime and things to maintain.
- Put the transformation logic in code that’s versioned and tested, and make every layer reproducible from the one before it.
- Apply governance per layer: raw layers may hold sensitive data that Silver masks or removes.
- Document each table: what it is, its grain, owner and freshness.
It’s a way to structure a data platform, not a product you install.