Implement a medallion architecture that survives contact with real data

Bronze is the landing zone and it has exactly one job: preserve what the source system sent, exactly as it sent it, with enough metadata to know when it arrived and where it came from. It is append-only. The moment you update Bronze in place you lose the ability to replay history, and replay is the whole reason the layer exists.

Silver is where data becomes trustworthy. Types are enforced, duplicates resolved, keys conformed across systems, and the messy realities — the same customer arriving from three systems with three spellings — are reconciled. Silver should be queryable by an analyst who knows the domain but not the source quirks.

Gold is shaped for consumption: aggregated, denormalized, named in business language, and built for specific consumers. A Gold table should be readable by someone who has never heard of your source systems. If your Gold tables still leak source-system column names, the Silver layer isn't finished.

The pattern originated in the Databricks ecosystem and is often described in Delta Lake terms, but nothing about it is vendor-specific. We have implemented the same layer contracts on Snowflake with dbt, on Microsoft Fabric, and on open Iceberg tables queried by several engines.

What matters is not the storage format but whether the boundaries are enforced. A medallion architecture on Delta with no tests and a mutable Bronze layer is worse than a well-governed three-schema warehouse, because it carries the vocabulary of rigor without the substance.

That said, the transformation tool does matter for how enforcement feels day to day. dbt and SQLMesh both give you version-controlled models, tests and lineage; they differ in how they handle incrementality, environments and change detection — a distinction worth understanding before you commit.

Sometimes not. For a small estate with a handful of clean sources and one consuming team, three layers can be ceremony that slows you down without buying much. Two layers, or one well-tested transformation step, may be the honest answer.

The layers earn their cost when you have multiple sources that disagree, multiple teams consuming the same entities, regulatory pressure to show where a number came from, or a history of quality incidents nobody could trace. If two or more of those apply, the structure pays for itself quickly.

We'll say so if your situation doesn't warrant it. Recommending architecture you don't need is how consultancies get a bad name.