Medallion Architecture: Bronze, Silver and Gold Layers

The medallion architecture explained: bronze, silver and gold layers in a lakehouse, what belongs in each, data quality gates and common mistakes.

Beginner 3 min read · Updated

The medallion architecture organises a lakehouse into three layers of increasing quality: bronze holds raw data as it arrived, silver holds cleaned, deduplicated and conformed data, and gold holds business-level tables ready for dashboards, applications and models. Each layer has a clear contract, so you always know where to fix a problem and what you can safely rebuild.

The three layers

cleanmodelSourcesCDC · events · filesBronzeraw · append-onlySilverclean · deduped · typedGoldmarts · aggregatesBIApps / APIsML features
Data flows from raw bronze through conformed silver to business-ready gold; each hop is a repeatable job with quality checks.
LayerContentsRulesWho reads it
BronzeSource data as received, plus load metadata (load time, source file, batch id)Append-only, minimal transformation, long retention, replayableData engineers, for reprocessing and debugging
SilverOne clean table per entity or event: typed columns, deduplicated, invalid rows quarantined, keys conformed across sourcesSchema enforced, quality checks, PII handled, history modelledData engineers, analysts, data scientists
GoldBusiness-level tables: facts and dimensions, aggregates, feature tables, product-specific martsOwned by a team, documented, tested against business rulesDashboards, applications, ML training, business users

What belongs in each layer

  • Bronze keeps what the source sent, including bad records and duplicates, because it's your recovery point. Store it in an open table format or as files partitioned by load date, and never edit it in place.
  • Silver is where most of the engineering lives: parse and type columns, drop exact duplicates, apply deduplication by key and event time, map source IDs to shared keys, split invalid rows into a quarantine table, and mask or tokenise sensitive fields.
  • Gold is shaped by questions: dimensional models, daily aggregates, one-big-table extracts for a dashboard, or feature tables for training.

Put data quality checks on each hop: shape and required fields into silver, uniqueness and referential integrity in silver, and business reconciliation (totals match the source system) into gold.

Why the layering helps

  • Reprocessing is cheap: a bug in silver logic is fixed by rerunning from bronze, without asking sources to resend data.
  • Ownership is clear: platform teams own bronze and silver standards; domain teams own their gold tables.
  • Blast radius is contained: a bad gold model breaks one mart, not every consumer.
  • Access control is simpler: most people only need gold; restricted raw data stays in bronze.

Common mistakes

  • Treating it as three copies of everything. Not every source needs a gold table; build gold for real consumers only.
  • Business logic in bronze. Once bronze is transformed, you've lost the raw record you'd need to fix that transformation.
  • Gold reading bronze directly. Skipping silver duplicates cleaning logic across marts and lets bad data leak.
  • No retention policy. Bronze grows forever unless you decide how long raw history is kept and how deletion requests (for example under privacy law) reach every layer.
  • Many tiny files. Streaming into bronze creates small files that slow every read; schedule compaction.

Medallion is a naming convention for an older idea (raw, staging and mart layers, or landing, curated and consumption zones). Teams use different names and sometimes more layers; what matters is that each layer has a contract and a single direction of flow.

Where it’s used

Lakehouse platforms

Bronze tables ingest CDC streams and files; silver conforms customers and orders across sources; gold serves finance and growth marts.

Event analytics

Raw app events land in bronze, silver sessionises and deduplicates them, and gold holds daily funnel and retention tables.

ML

Feature pipelines read silver tables directly for detail, while gold holds curated training datasets with lineage.

In an interview

  • Keep bronze raw and replayable, and say how a backfill reruns silver and gold.
  • Put deduplication, typing and PII handling in silver.
  • Name the quality gate between each layer.
  • Build gold only for named consumers, with owners.

Further reading