Medallion architecture fails when the layers stop behaving like a pipeline. Microsoft’s Fabric Data Warehouse blog on 3 September 2026 put it bluntly: bronze starts cleaning, silver starts serving dashboards, gold starts compensating for upstream quality. The names bronze, silver, and gold are not the problem. The jobs getting mixed is the problem.
What each layer is for
Bronze lands what arrived, with metadata, in batches. It should preserve the source’s mess. Silver validates, conforms, and flags. Gold exposes trusted fields for a use case. If you clean in bronze, you cannot replay a source dispute. If you serve dashboards from silver, you will be afraid to change silver. If gold patches quality, nobody will fix producers.
Citi Global Wealth, as described by Computer Weekly on 8 September 2026, used bronze for per-source models, silver as a unified data model, and gold for specific uses. Business stakeholders helped write the silver taxonomy. That split is the pattern worth copying, not the bank’s vendor list.
Anti-pattern one: bronze as a spa
Teams “help” bronze by dropping nulls, recoding statuses, and aligning time zones. Then a source system changes a code and nobody can prove what arrived. Keep bronze ugly. Put the cleaning rules in silver where they have tests.
Microsoft’s checklist in that September 2026 post says bronze should batch ingest with metadata and preserve what arrived. Tiny row-by-row writes are another bronze failure. Warehouses want set-based loads.
Anti-pattern two: silver as a BI tool
Silver is a language, not a dashboard store. The moment executives live in silver tables, you cannot rebuild them. Gold exists so silver can stay rerunnable. If an analyst needs a pretty grain, make a gold mart. Do not add a column to silver “just this once.”
Anti-pattern three: gold as a hospital
Gold that coalesces nulls, hard-codes a customer name, or averages over broken keys is hiding a silver bug. The Fabric post says gold should expose trusted fields only, not compensate for upstream quality. Send the defect back.
How to inspect your own pipeline this week
Pick one KPI. Ask which layer last changed its value. If the answer is a gold CASE statement, you have the hospital pattern. If the answer is a bronze notebook, you have the spa. If nobody knows, you do not have medallion. You have a folder of SQL.
Google Cloud’s 18 September 2026 lakehouse note on cross-cloud caching is about movement cost. It does not excuse mixed layers. Caching a confused gold mart only makes the wrong number cheaper.
Fivetran and dbt merged in late 2025. Combined tooling still cannot decide your layer jobs. You can. Write the three jobs on a wiki page and reject pull requests that violate them.
Frequently asked questions
What is the main medallion anti-pattern?
Mixing jobs: cleaning in bronze, serving dashboards from silver, patching quality in gold.
Who said this in 2026?
Microsoft’s Fabric Data Warehouse blog on 3 September 2026, in a series on medallion with Fabric.
Should bronze ever clean PII?
Tokenize or restrict access, but keep a recoverable raw landing. Do not make bronze the place business rules hide.
Can silver be queried by people?
For debugging and contracts, yes. For executive dashboards, no. Those belong on gold.
How do I fix gold that patches quality?
Move the rule to silver with a test, or to a data contract on the producer. Then delete the gold patch.
Is medallion mandatory?
No. Honest layers are mandatory. If you have one warehouse, still separate land, conform, and serve.
Book a demo to see a managed data platform running on your own sources
LensAIght
