Architecture on Databricks is mostly about how data is organised as it moves from raw ingestion to business-ready tables. The pages here describe the medallion architecture, the layered convention used by most Databricks lakehouses, layer by layer.
Architectural patterns for organising data on Databricks. Everything here currently sits under the medallion model — raw landing, cleaned, and business-ready layers.
Bronze tables hold data as it arrived, with ingestion metadata, so any later step can be replayed. Silver tables deduplicate, fix types, apply CDC and join reference data to produce a trustworthy, conformed view of each entity. Gold tables are shaped for a specific use: star schemas for BI, feature tables for machine learning, or aggregates for an application. All three layers are Delta tables, typically governed in Unity Catalog with one schema or catalog per layer.
| Layer | Contents | Quality | Main consumers |
|---|---|---|---|
| Bronze | Raw records, ingestion time, source file | As received | Data engineers, replay |
| Silver | Validated, deduplicated, current-state and history tables | Enforced schema and expectations | Analysts, data scientists |
| Gold | Dimensional models, KPIs, features | Business-ready | BI dashboards, apps, ML |
The pattern is a convention, not a rule. Small projects sometimes merge silver and gold; regulated ones may add a layer for masked or anonymised data. Partitioning and clustering choices differ per layer, which is the subject of the partitioning code page below.