Is your data platform helping people make better decisions – or just giving them faster access to the same old mess?
There is no doubt Microsoft knows what it is doing when it comes to enterprise data and analytics. You do not stay a Leader in the Gartner Magic Quadrant for Analytics and Business Intelligence Platforms for 18 years without serious credentials, sustained innovation, and the trust of enterprise customers.
Microsoft Fabric is the next chapter in that story. It brings together data engineering, analytics, real-time intelligence, governance, and AI on a single platform. That is a powerful promise. But as many teams quickly discover, the platform itself does not create trusted data. The architecture does.
Without the right design, Fabric can become another version of the fragmented environment it was meant to replace. And before long, people are still asking the same uncomfortable question:
“Can we trust this number?”
Most enterprises are converging on the same solution: the medallion architecture. This is the pattern Microsoft formally recommends for Fabric, and for good reason – it aligns perfectly with how Fabric stores, governs, and serves data through OneLake.
This guide looks at how medallion architecture works in Microsoft Fabric, where it adds the most value, and the pitfalls to avoid.
What Medallion Architecture Actually Means
At its core, medallion architecture organises data into three progressive quality tiers: Bronze, Silver, and Gold.
Each tier represents a higher level of refinement, moving from raw, unfiltered data to business-aligned, AI-ready insights.
In Microsoft Fabric, this design sits atop OneLake, the unified, tenant-wide data lake. Because every layer uses Delta Lake as the default storage format, all tiers share a common, open foundation. This eliminates proprietary silos and prevents data duplication across different engines.
Bronze: Your Immutable Source of Truth
Bronze is where raw data lands, preserved in a state as close to the source format as possible. A few key principles apply here:
- Ingest, Don’t Transform: Bronze should capture data in its original schema with minimal processing beyond basic validation. When the source is a relational database, Delta tables are the ideal default.
- Leverage Shortcuts and Mirroring: Fabric’s OneLake shortcuts allow you to virtualize data from various sources without moving it, reducing both cost and latency. Mirroring provides a low-code way to handle operational databases with real-time change data capture (CDC).
- Design for Replay: By keeping a raw history in Bronze, you can reprocess Silver and Gold whenever logic changes without having to re-ingest data from source systems.
Silver: Cleansed, Conformed, and Enterprise-Ready
Silver is where data becomes useful. In this layer, you deduplicate, standardize, and enforce data types. It’s also where you join data across disparate sources into a coherent model.
Typical Silver workloads include:
- Spark Notebooks: (PySpark, Spark SQL, or Scala) for large-scale transformations. A best practice is to ingest each source separately into Bronze and combine them in Silver to keep pipelines modular.
- Dataflows Gen2: Ideal for lower-code transformations where the team prefers a Power Query experience.
- Embedded Quality Checks: Data quality logic should be baked into the pipeline, ensuring failed records are quarantined rather than silently dropped.
Silver should be the primary source for data scientists and advanced analysts who need clean, granular data without the specific business logics applied in Gold.
Gold: Business-Ready and Consumption-Optimized
Gold is the layer the business actually sees. It is modelled for performance, aligned to specific domains, and designed to answer critical business questions.
Fabric supports two common architectural patterns for the Gold layer:
- Lakehouse Gold: Each layer sits in a Lakehouse, and users query data via the SQL analytics endpoint. This is best if your team is Spark- or Python-native.
- Warehouse Gold: Bronze and Silver are Lakehouses, while Gold is a Fabric Warehouse. This is ideal when your data is highly structured and your consumers/developers are SQL-first.
The deciding factor isn’t technical superiority—both patterns read and write the same Delta Parquet files in OneLake. Instead, choose the pattern that best fits your team’s existing skillset.
Design Decisions That Separate Good from Great
Getting the layers right is only half the battle. These decisions often define the long-term success of a Fabric implementation:
- One Workspace Per Layer: To maintain enterprise-scale management, Microsoft recommends creating each Lakehouse in its own workspace. This simplifies permissions, deployment pipelines, and capacity management.
- Default to DirectLake for Power BI: DirectLake reads Delta Parquet files directly from OneLake into the in-memory engine. This provides “Import” performance with “Live” freshness—meaning no scheduled refreshes are required.
- Adopt CI/CD Early: The fabric-cicd Python library is now the officially supported tool for deployment automation. Version-control your notebooks, pipelines, and semantic models from day one.
- Use Schemas to Separate Domains: If you prefer fewer Lakehouses, use schemas to cleanly partition domains like Sales, Finance, or Operations within a single environment.
- Bake in Governance Early: Microsoft Purview integrates natively with Fabric. Sensitivity labels, lineage, and Data Loss Prevention (DLP) should be configured as you build, not as an afterthought.
Common Pitfalls to Avoid
- Over-Engineering Bronze: Heavy transformations in Bronze defeat its purpose as an immutable landing zone and make reprocessing difficult.
- Bloating Gold: Resist the urge to load unfiltered datasets into Gold. It should contain curated, modelled data—not raw dumps with a new label.
- Treating Medallion as Dogma: The three-layer pattern is a guide, not a rigid rulebook. There are rare use cases where a simplified two-layer approach may be more efficient.
- Underestimating Capacity Planning: Fabric’s elastic scaling isn’t a license for poor workload management. Monitor capacity consumption closely and right-size your capacity to avoid performance bottlenecks.
Architecture by Design, Not by Default
Medallion architecture is most beneficial when an organization deals with high scale, frequently changing source schemas, or distributed data ownership across multiple teams or domains. In these more complex environments, the extra physical layers serve as necessary boundaries to manage scale and shifting requirements effectively.
However, we often find Fabric environments that are overbuilt and would benefit from a simpler approach. Instead of materializing redundant Bronze, Silver, and Gold tables that create unnecessary overhead, a simplified model that utilises just two primary zones: a Landing Zone for ingestion and a Curated Layer for consumption is more suitable for environments where the data consists of stable, structured inputs and a well-defined schema that rarely changes. This streamlined strategy not only reduces operational costs but also ensures that complex layers are only introduced when specifically justified by business needs, resulting in a cleaner, more maintainable data lineage.
Key Takeaways
- Choose the Right Architecture: Medallion architecture is the recommended design for Fabric. However, do assess if that’s the right design for your organisation.
- Focus on Purpose: Use Bronze for raw history, Silver for clean enterprise data, and Gold for business-ready insights.
- Skillset Matters: Choose between Lakehouse and Warehouse for Gold based on what your engineers can maintain best.
- Prioritise Operations: Treat workspace design, DirectLake, CI/CD, and Purview governance as first-class citizens.
The architecture you choose today determines how quickly you can embed AI and autonomous agents into your analytics tomorrow.
How Skillfield Can Help
At Skillfield, we help medium and large enterprises design and implement modern data platforms on Microsoft Fabric – from initial strategy to production-grade deployment. Our architects work across Lakehouses, Warehouses, Real-Time Intelligence, Purview, and Power BI to deliver unified analytics environments that scale.
Whether you’re migrating from a legacy estate or standing up Fabric for the first time, we bring a governance-first mindset and pragmatic delivery experience to every engagement.
Ready to build a Fabric platform your business can trust? Talk to a Fabric expert at Skillfield about your medallion architecture roadmap, implementation, or health check. Visit www.skillfield.com.au to start the conversation.
Author: Anton Fernando
Further Reading:
https://skillfield.com.au/blog/why-most-ai-projects-fail/
https://skillfield.com.au/blog/bring-reporting-home-smarter-faster-onshore-solutions/
