The Medallion Architecture in Azure Synapse Analytics: A Strategic Guide for Data Leaders
Raw to trusted to curated, with a clear path to Fabric.
Medallion architecture in Azure Synapse is a practical answer to a familiar problem: data that nobody fully trusts. By moving data through bronze, silver and gold layers, you separate raw capture from validated data and from the curated tables that people actually use for decisions.
This guide explains each layer as it is built on Azure Synapse, the business case, the common pitfalls, and how to move the same design to Microsoft Fabric. It is written for data leaders deciding how to structure a platform that will outlast any one tool.
What Is the Medallion Architecture?
The medallion architecture is a design pattern for organising data in a lake. It has three layers: bronze for raw data, silver for enriched and validated data, and gold for curated data. Each layer indicates the quality of the data it holds, with higher layers carrying higher quality.
The key idea is that raw data is never treated as reporting-ready. The original copy is preserved as a source of truth, while pipelines of validation and transformation prepare data for analytics. A less obvious benefit is recovery: when a transformation is wrong, you rebuild silver and gold from bronze instead of re-extracting from source systems.
The Bronze Layer: Raw Data Ingestion and Full Fidelity
Bronze stores data as received from source systems, with its full history and ingestion metadata. In Azure Synapse, bronze usually lives in Azure Data Lake Storage Gen2, loaded by Synapse pipelines, Azure Data Factory or Event Hubs for streaming sources. Keep the original format where you can.
Why Raw Fidelity Matters for Governance
A faithful bronze layer supports audit, lineage and reprocessing. When a regulator or auditor asks where a number came from, you can trace it back to the data as it arrived.
Retention: Keep What You Must, Then Delete
Bronze should not be purged casually, but it should not be kept forever either. Set a retention policy that matches your regulations and your reprocessing needs. Azure Blob Storage supports immutable (WORM) storage with time-based retention policies and legal holds, which prevent modification or deletion during the retention period. Container-level policies work with accounts that have a hierarchical namespace, which Data Lake Storage Gen2 uses, while version-level policies are not yet supported there.
The Silver Layer: Validated, Conformed and Analytics-Ready
Silver is where data becomes trustworthy. Typical work includes deduplication, type casting, referential checks and business rules, producing consistent entities such as customer, product and order. In Synapse this is usually done with Spark pools or mapping data flows and stored as Delta or Parquet tables.
Silver as the Foundation for Self-Service
Because silver is conformed but not aggregated, analysts and data scientists can reuse it for many questions. Keep pre-aggregations out of silver. They belong in gold, where they serve a defined consumer.
The Gold Layer: Curated Business Intelligence and Decision Data
Gold holds purpose-built, consumer-shaped data: aggregated measures, dimensional models and domain data products. In Synapse it is often materialised as dedicated SQL pool tables or external tables, and served to Power BI and SQL clients.
Gold Layer Examples Across Industries
A retailer might publish sales by store, product and week. A financial services firm might publish exposure by counterparty and day. A manufacturer might publish yield and downtime by line and shift. In each case, gold answers a defined set of business questions.
Analysis Services and Power BI Semantic Models
Azure Analysis Services can still consume gold, and Microsoft states there are currently no plans to deprecate it. However, Microsoft is focusing investment on Power BI semantic models in Fabric for enterprise data modelling, so new models usually belong there.
Bronze vs Silver vs Gold: A Practical Comparison
| Dimension | Bronze | Silver | Gold |
|---|---|---|---|
| State | As received | Validated and conformed | Curated and aggregated |
| Purpose | Preserve full fidelity and history | Create a reusable, trusted base | Serve specific business needs |
| Typical consumers | Pipelines, audit, reprocessing | Analysts, data scientists, other pipelines | Reports, dashboards, applications |
| Transformations | Minimal | Cleansing, typing, deduplication, rules | Aggregation, business logic, shaping |
| Schema | Flexible | Enforced | Fixed for the consumer |
| Format | Original format, Parquet or Delta | Delta or Parquet | Delta, warehouse tables or semantic models |
| Access | Restricted | Broad for analysts | Based on consumer need and least privilege |
| Governance value | Audit trail and lineage | Quality evidence | Certified, trusted outputs |
The Business Case for Medallion Architecture in Azure Synapse
The case rests on three returns, none of which this guide quantifies, because they depend on your estate. First, lower incident cost: when data is wrong, you fix one layer and rebuild, instead of chasing copies. Second, faster new use cases: a trusted silver layer means a new report starts from conformed data, not from raw extracts. Third, a stronger governance posture: access rules and quality gates apply per layer, so consumers see only what suits their role.
To put numbers on it, baseline incident hours, time to first report for a new use case, and the number of copies of key tables before you start.
Common Implementation Pitfalls and How to Avoid Them
Defining Layer Boundaries Too Loosely
If teams put aggregates in silver or raw fixes in gold, the layers stop meaning anything. Write down what each layer may contain, and enforce it in review.
Silver Lake Bloat
Silver grows quickly when every team adds its own variations. Prefer a small number of shared, conformed entities, and retire unused tables.
Underinvesting in Bronze Retention
Without a retention policy, bronze either grows without limit or is cleaned up ad hoc. Agree the rule with compliance, apply it with lifecycle management, and document it.
- Bronze preserves raw data, silver makes it trustworthy, gold serves the business.
- Set retention deliberately: keep what regulation requires, then delete.
- Keep layer boundaries written down and enforced.
- Synapse is not retired, but Microsoft positions Fabric as the platform for new innovation.
- The same medallion design carries into Fabric, often with shortcuts instead of copies.
Moving the Medallion Architecture in Azure Synapse to Fabric
Microsoft has not announced a retirement of Azure Synapse Analytics or a required migration date. It does describe Microsoft Fabric as its strategic platform for new analytics innovation and encourages customers to evaluate it for new investments and long-term modernisation. That makes the question one of timing and value, not urgency.
The good news is that the design carries over. Fabric describes medallion architecture as its recommended design approach, implemented with lakehouses in OneLake, either one per layer or lakehouses for bronze and silver with a warehouse for gold. Microsoft recommends separate workspaces per layer for governance and control.
| Component | In Azure Synapse | In Microsoft Fabric |
|---|---|---|
| Storage | ADLS Gen2 containers per layer | OneLake shortcuts to the same ADLS Gen2 paths (zero copy), or data copied into OneLake |
| Spark transformations | Synapse Spark pools and notebooks | Fabric notebooks and environments; the Spark Migration Assistant (preview) moves supported items, then you refactor code |
| Ingestion and orchestration | Synapse pipelines and Azure Data Factory | Fabric Data Factory pipelines; mapping data flows map to Dataflow Gen2 |
| Gold: SQL serving | Dedicated SQL pool tables | Fabric Data Warehouse, moved with the Migration Assistant for Data Warehouse |
| Gold: semantic models | Power BI import or Analysis Services models | Power BI semantic models, using Direct Lake on gold Delta tables |
| Governance | Layer-specific access and Purview | OneLake security, catalog endorsement and lineage across layers |
Start with the data, not the code
OneLake shortcuts to existing ADLS Gen2 paths are the recommended zero-copy approach. Delta tables placed in the Tables section of a lakehouse register automatically, and become visible to the SQL analytics endpoint and Power BI.
Expect refactoring in Spark and pipelines
The Spark Migration Assistant, currently in preview, moves notebooks, Spark job definitions, pools and lake database metadata, but it does not complete the job. Spark configurations and custom libraries are not migrated, Synapse workspaces under a VNet cannot be migrated with it, and code that uses mssparkutils needs to move to notebookutils. Microsoft also does not backport new Fabric pipeline features into Synapse pipelines, so the two evolve separately.
Move the gold layer last, and validate it
Run old and new gold layers in parallel, compare outputs, then repoint Power BI reports. For the warehouse, plan for T-SQL differences. Our two-tier medallion architecture post shows a simpler layer design for Fabric, and the Microsoft Fabric migration checklist keeps each wave consistent. For the platform decision, see Azure Synapse vs Microsoft Fabric.
Medallion Architecture in Azure Synapse FAQ
What is the medallion architecture?
It is a design pattern that organises data in three layers: bronze for raw data, silver for validated and conformed data, and gold for curated data shaped for consumers. Data quality improves at each step, and the raw copy is kept as a source of truth.
Can I run a medallion architecture in Azure Synapse?
Yes. A common pattern stores bronze, silver and gold in Azure Data Lake Storage Gen2, uses Synapse pipelines and Spark for movement and transformation, and serves gold through dedicated SQL pools or Power BI. Whether Synapse is the right long-term home is a separate decision, covered in the Fabric section.
Is Azure Synapse Analytics being retired?
Microsoft has not announced a retirement of Azure Synapse Analytics or a required migration date. It does describe Microsoft Fabric as its strategic platform for new analytics innovation and encourages customers to evaluate Fabric for new investments and long-term modernisation.
Should bronze data ever be deleted?
Yes, when retention rules allow it. Keep bronze for as long as your regulations and reprocessing needs require, then delete it. For regulated data, Azure Blob Storage supports immutability policies such as time-based retention and legal holds.
Do I need to rebuild the layers to move to Fabric?
Not usually. Fabric recommends the same medallion pattern, and OneLake shortcuts can point to your existing Delta data in ADLS Gen2 without copying it. Code, SQL and pipelines still need review and refactoring.
Is Azure Analysis Services still a valid gold consumer?
It works, and Microsoft states there are currently no plans to deprecate it. Microsoft's investment is focused on Power BI semantic models in Fabric, so new gold-layer models are usually built there.
Next Steps: Building Your Medallion Architecture on Azure
Treat medallion as a default starting point, adopt it incrementally by business domain, and decide on Fabric when the value is clear. For delivery support, see our data engineering practice, our Azure Synapse consulting and Microsoft Fabric migration services, or talk to our team.