What is Medallion Architecture? The Bronze-Silver-Gold Pattern, Defined

Medallion architecture is a layered data design pattern (bronze, silver, gold) where raw data lands in bronze, is cleaned and conformed in silver, and is aggregated for analytical consumption in gold, popularized by Databricks for lakehouse deployments.

Medallion architecture is a layered data design pattern, popularized by Databricks for lakehouse deployments, where raw data lands in a bronze layer, is cleaned and conformed in a silver layer, and is aggregated for analytical consumption in a gold layer. Each layer is typically materialized as Delta or Iceberg tables in the lakehouse.

What medallion architecture actually is

Bronze, silver, gold: three layers with distinct jobs

Each layer is a Delta or Iceberg table in the lakehouse, so every stage stays queryable and replayable.

• Bronze: append-only raw ingestion, schema mirrors the source, built for faithful capture and replay
• Silver: deduplicated, type-normalized, schema-conformed canonical entities as the enterprise source of truth
• Gold: dimensional models, aggregated metric tables and feature tables served with no further transformation

Bronze to silver is usually Spark notebooks. Silver to gold is usually dbt or SQL. The split gives ingestion, conformance and analytical modelling their own owners.

Source: Thinklytics data platform practice, 2026.

Three layers with distinct jobs:

• Bronze (raw): append-only ingestion of source data with minimal transformation. Schema mirrors source systems. Job: capture data faithfully and enable replay.
• Silver (cleaned): deduplicated, type-normalized, schema-conformed canonical view of each entity. Job: provide the enterprise source of truth for downstream consumers.
• Gold (aggregated): dimensional models, aggregated metric tables, and feature tables tuned for analytical and ML consumption. Job: serve dashboards, BI tools, and model training without further transformation.

Each layer is a Delta or Iceberg table in the lakehouse. Transformations between layers are usually Spark notebooks (bronze to silver), and dbt or SQL (silver to gold). The layers create natural separation of concerns for ingestion, conformance, and analytical modeling work.

What people confuse it with

Medallion and classic warehouse layering

• Traditional warehouse. ODS, DW, DM. Operational data store, warehouse, then data mart. The same separation of concerns, decades older.
• Medallion. Bronze, silver, gold. The lakehouse-era version, with open table formats, separate Spark and SQL compute, and version control on the transformations.

Layers do not need separate physical Lakehouses. They can be folder prefixes, separate Delta tables inside one Lakehouse, or separate Lakehouses, and the right granularity follows access control and team ownership.

Source: Thinklytics data platform practice, 2026.

• "Medallion is just bronze, silver, gold rebranding of ODS, DW, DM." Functionally similar with lakehouse-era ergonomics (open table formats, separate Spark and SQL compute, version control on transformations).
• "Medallion is mandatory for a lakehouse." Not true. Flat layouts work for prototypes. For production tenants with 2+ workloads, medallion's separation pays back quickly.
• "Each layer must be a separate physical Lakehouse or Workspace." Wrong. Layers can be folder prefixes, separate Delta tables within a single Lakehouse, or separate Lakehouses. The right granularity depends on access-control and team ownership patterns.

When medallion architecture matters

When medallion layering earns its keep

Three signals the separation pays back, and three ways teams make it cosmetic.

• More than 2 production workloads on the lakehouse. Shared raw-data ingestion has become a bottleneck.
• Multiple downstream consumers read the same entities. BI, ML and operational systems all need one canonical customer or account table.
• Schemas evolve and replay is needed. Rebuild gold from silver, rebuild silver from bronze, without going back to the source systems.
• A single dashboard against a single table. The layering overhead is not justified at that size.
• Medallion treated as documentation only. The layer boundaries are drawn but never enforced in practice.
• Pipelines that cross several layers in one run. The architectural separation becomes cosmetic rather than real.

Medallion is a recommended pattern, not a requirement. Flat layouts work for prototypes, and for any tenant with more than 2 production workloads the separation pays back quickly in maintainability.

Source: Thinklytics data platform practice, 2026.

Medallion matters when:

• The lakehouse has more than 2 production workloads and shared raw-data ingestion is a bottleneck.
• Multiple downstream consumers (BI, ML, operational systems) read from the same canonical entities.
• The team needs replay capability (rebuild gold from silver, rebuild silver from bronze) when schemas evolve.

When medallion does not help

Medallion does not help when:

• The workload is a single dashboard against a single table, where the layering overhead is not justified.
• The team treats medallion as documentation only, without enforcing the layer boundaries in practice.
• The transformations cross too many layers in a single pipeline, making the architectural separation cosmetic.

How Thinklytics works on medallion architecture

Medallion design is part of every lakehouse engagement, with attention to per-layer SLAs, transformation tooling, and schema-evolution patterns that hold up at scale. See Microsoft Fabric data engineering.

Frequently asked questions

What is medallion architecture in one sentence?

Medallion architecture is a layered data design pattern, popularized by Databricks, where raw data lands in a bronze layer, is cleaned and conformed in a silver layer, and is aggregated and modeled for analytical consumption in a gold layer, with each layer materialized as Delta or Iceberg tables in a lakehouse.

What is in the bronze layer?

Raw, append-only data from source systems, with minimal transformation. The bronze layer's job is to capture source data faithfully and provide replay capability. Schema is often loose or source-mirrored. Typical contents: source-system tables ingested as-is, CDC change feeds, raw API payloads, file dumps.

What is in the silver layer?

Cleaned and conformed data with consistent schema, deduplication, type normalization, and standardized naming. The silver layer is the enterprise-trusted canonical view of each entity. Typical contents: cleaned customer, account, transaction, and event tables that downstream consumers trust as the source of truth.

What is in the gold layer?

Aggregated, business-ready data optimized for analytical consumption. The gold layer is what dashboards, BI tools, and ML training pipelines actually consume. Typical contents: dimensional models (star schemas), aggregated metric tables, feature tables for ML, and curated views per consumer use case.

Is medallion architecture the same as data warehouse layering?

Closely related. Traditional data warehouses had ODS (operational data store), DW (warehouse), and DM (data mart) layers. Medallion is the lakehouse-era version with similar separation of concerns but with open table formats and Spark-or-SQL compute.

Is medallion architecture mandatory in a lakehouse?

No. It is the recommended pattern from Databricks and a common pattern in Microsoft Fabric, but flat layouts work for small prototypes. For any tenant with more than 2 production workloads, medallion's separation of concerns pays back quickly in maintainability.

How do you implement medallion in Microsoft Fabric?

In OneLake, the typical pattern is one Lakehouse per domain with bronze, silver, gold as folder prefixes or as separate Delta tables within the Lakehouse. The Fabric Warehouse usually serves the gold layer where Power BI Direct Lake mode reads from. See Microsoft Fabric data engineering.

How does Thinklytics work on medallion architecture?

Medallion design is part of every lakehouse engagement we run, with attention to per-layer SLAs, transformation tooling (dbt for gold, Spark notebooks for bronze and silver), and the schema-evolution patterns that hold up at scale. See Microsoft Fabric data engineering.