Medallion Architecture
The medallion architecture organizes a data lake into bronze, silver, and gold layers, refining raw data into curated, consumption-ready tables in stages.
Refinement in layers
The medallion architecture structures a data lake as a progression of quality tiers, conventionally named bronze, silver, and gold. Each layer takes the previous one and refines it, so raw data becomes progressively cleaner, more integrated, and more usable. The named layers make responsibilities explicit and give every table a clear place in the pipeline.
Bronze: raw ingestion
The bronze layer holds data as it arrived, appended with minimal or no transformation, keeping source fidelity and full history. Its purpose is to be a faithful, replayable record: if a downstream logic error is found, you can reprocess from bronze rather than re-ingest from the source, which may no longer hold the data. Bronze prioritizes completeness over cleanliness.
Silver: cleaned and conformed
The silver layer applies cleaning, deduplication, type enforcement, and conforming of records from multiple sources into consistent, validated tables. This is where data-quality expectations are enforced and where schemas are standardized. Silver tables are trustworthy and queryable but still relatively normalized and general-purpose rather than shaped for a specific consumer.
Gold: business-ready
The gold layer holds aggregated, denormalized tables shaped for specific consumption: dashboards, reports, feature sets, and model inputs. Gold tables answer particular questions efficiently and often follow dimensional-modeling patterns. Because they are derived, they can be rebuilt from silver whenever definitions change.
Why the layering helps
- Failures are isolated to the layer where they occur
- Reprocessing starts from the last good layer, not the original source
- Each layer has a clear owner, contract, and quality bar
- Consumers read gold without touching messy raw data
The architecture pairs naturally with orchestration, where each layer transition is a pipeline stage, and with quality validation at the bronze-to-silver boundary. See workflow orchestration, data-quality validation, and the lakehouse.