SHORT ANSWER
Medallion architecture organises data in a lakehouse into three layers of increasing quality: bronze for raw data as ingested, silver for cleaned and conformed data, and gold for business-ready tables and aggregates. You rebuild each layer from the one below, which makes pipelines easier to debug, audit and reprocess without going back to source systems such as ERP or CRM.
Medallion architecture is a design pattern that organises a data lakehouse into layers named after medals: bronze, silver and gold. Data moves from raw to refined as it passes through the layers, and each layer has a clear purpose, quality level and audience. Databricks popularised the pattern, and Microsoft recommends it for Fabric too.
How does medallion architecture work?
Bronze (raw): data lands as it arrives from source systems, files, APIs or streams, often via change data capture. You keep everything and record ingestion time and source.
Silver (cleaned): data is deduplicated, validated, typed and joined into consistent entities such as customers, products and orders.
Gold (business-ready): data is modelled into facts, dimensions, KPIs and ML features, ready for BI tools and applications.
Most teams use Delta Lake tables at every layer, so each step is transactional and you can query past versions.
Why does it matter for enterprises?
A clear layer structure gives your teams a shared vocabulary and makes it obvious where to fix a problem. Because bronze keeps the raw data, you can change business logic in silver or gold and rebuild without re-extracting from ERP or CRM systems. You can also tier access: only data engineers see bronze, while analysts work in gold. That keeps personal and sensitive data easier to govern.
The three layers at a glance
| Bronze | Silver | Gold |
|---|---|---|---|
Content | Raw source data | Cleaned, conformed entities | Business aggregates and models |
Typical users | Data engineers | Engineers, data scientists | Analysts, BI, applications |
Quality checks | Minimal | Validation and deduplication | Business rules and KPIs |
Schema | As source | Standardised | Modelled for use |
Access rules are where layers earn their keep. On a HIPAA-aligned Databricks platform we built on Azure for a nonprofit human-services organization, we used Unity Catalog for central governance, lineage and access control inside a fully private network. With sensitive health data, you want to grant access by layer and role, not table by table, and a layered design makes those rules simple to state.
A few rules keep the pattern useful. Bronze is append-only and nobody edits it by hand. Silver holds one agreed version of each business entity, with personal data classified. Gold tables serve specific consumers, such as a Power BI semantic model, and you document them as data products. Without these rules, layers blur and teams end up querying bronze directly.
How RUBICON helps with lakehouse design
Our data engineering team builds governed lakehouse platforms on Azure and Databricks. If you’re designing layers for a new platform, our architects can review your sources and reporting needs with you.
Related terms
Frequently asked questions
What is the difference between bronze, silver and gold layers?
Bronze holds raw data exactly as received from source systems, with ingestion metadata. Silver holds cleaned, deduplicated and conformed data, with consistent types and keys. Gold holds business-level tables such as facts, dimensions, KPIs and features, shaped for reports, semantic models and machine learning. Quality and trust increase at each layer.
Is medallion architecture only for Databricks?
No. Databricks popularised the name, but the pattern works on any lakehouse or warehouse. Microsoft Fabric documentation recommends it for OneLake, and similar layered designs (often called raw, staging and marts) are common in Snowflake and dbt projects. The idea of step-by-step refinement doesn't depend on the platform.
Do I always need three layers?
Not always. Some teams add a platinum or serving layer, while small projects sometimes merge silver and gold. The number of layers matters less than clear rules for what each layer contains, who can access it and how data moves between layers. Three layers is a sensible default for most enterprise platforms.
Does medallion architecture increase storage costs?
It stores data more than once, but object storage is cheap compared with compute, and you can compress bronze data or move it to cooler storage tiers. Being able to reprocess from bronze without going back to source systems usually outweighs the extra storage cost, especially when sources are slow or you keep data for audits.
Related case study
HIPAA Aligned Data Platform on Azure & Databricks
How RUBICON Delivered a Secure, Scalable Foundation for Healthcare Data, Analytics and Machine Learning
More resources
