Medallion Architecture

A layered data-design pattern that refines data through successive named tiers — conventionally bronze (raw landing) → silver (cleaned/conformed) → gold (business-level, modeled) — so each layer has a clear contract with the next.

Grounded 2026-06-26 by Databricks — What is Medallion Architecture (fetched under the auto-ground convention). Databricks is the pattern's popularizer — its doc defines the layers as data-quality tiers and explicitly describes bronze as where source data lands, which is the literal basis for the Lands vs Lives distinction. (The doc names no individual coiner.)

Overview

The medallion (or "multi-hop") pattern organizes a pipeline into quality tiers rather than one monolithic transform. Databricks' canonical framing (Databricks — What is Medallion Architecture) is that the layers exist to "incrementally and progressively improve the structure and quality of data as it flows through each layer." Raw data lands in bronze (optimized for CDC, historical archive, lineage, auditability; minimal validation); cleaning, conforming, and de-duplication happen in silver (at least one validated, non-aggregated record per source row); business modeling and de-normalized, read-optimized serving shapes live in gold. The value is that each hop has an explicit job and an explicit hand-off, which is what makes the lineage legible — and the whole stack is pitched as "a single source of truth" (System of Record).

A view-only medallion is a variant the source (Working Heuristics — Data, Matching, Product, Org, Tooling) calls out: the upper layers are views over the one physical base table rather than separate materializations. In that case the distinction between where data lands (the single physical base table) and where it lives (the view layer that does the modeling) becomes the key thing to reason about — see Pipeline Legibility and the lands-vs-lives distinction.

Details

The pattern presupposes the disciplines this wiki files separately:

• Data Contract — the hand-off between layers is exactly where derivation (timezone/unit/format) should happen once, so downstream tiers inherit correct values.
• One Fact Per Grain — gold-layer facts and dimensions earn their quality tier by respecting grain.
• Pipeline Legibility — naming each layer for what it does (and which direction data flows through it) is what keeps a multi-hop pipeline readable.

Connections

• Pipeline Legibility — locating where data lands vs lives in a view-only medallion is the headline application of reading a pipeline honestly.
• Data Contract — the inter-layer boundary is the natural home for the contract.
• One Fact Per Grain — the modeling discipline the gold layer is supposed to embody.

Sources

• Working Heuristics — Data, Matching, Product, Org, Tooling — supplies the view-only-medallion "lands vs lives" framing.
• Databricks — What is Medallion Architecture — the canonical bronze/silver/gold definition (quality tiers; bronze as the landing layer; "single source of truth").