Medallion Architecture Cheat Sheet

Updated 2026-05-28

Next Topic: Prefect Data Orchestration Cheat Sheet

🧠Study flashcards on this topic164 cards · spaced repetition→

The Medallion Architecture is a layered data design pattern — typically Bronze → Silver → Gold — that structures a data lakehouse by progressively refining raw data into trusted, analytics-ready assets. First popularized by Databricks with Delta Lake, the pattern is now format-agnostic and runs on Apache Iceberg, Apache Hudi, and Apache Paimon, making it a universal blueprint for modern data platforms regardless of cloud or vendor. The key mental model is separation of concerns by transformation stage: Bronze preserves raw fidelity, Silver enforces integrity and conformity, and Gold serves specific business consumers — and the boundary between Silver and Gold is best tested by whether a transformation requires domain knowledge (if yes, it belongs in Gold, not Silver).

What This Cheat Sheet Covers

This topic spans 23 focused tables and 193 indexed concepts, 164 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core Architecture LayersTable 2: Bronze Layer Ingestion PatternsTable 3: Bronze Layer Design PrinciplesTable 4: Silver Layer Transformation StandardsTable 5: Silver Layer Quality EnforcementTable 6: Gold Layer Business AggregationsTable 7: Data Vault and Dimensional Modeling IntegrationTable 8: Semantic LayerTable 9: Schema Evolution HandlingTable 10: Open Table FormatsTable 11: Performance Optimization TechniquesTable 12: Data Quality Layers AssignmentTable 13: Multi-Hop Architecture PatternsTable 14: Tooling and FrameworksTable 15: Platform-Specific ImplementationsTable 16: Access Control and GovernanceTable 17: Common Anti-Patterns and MistakesTable 18: Optimization Best PracticesTable 19: Naming ConventionsTable 20: Testing and Validation PatternsTable 21: Cost Optimization StrategiesTable 22: Monitoring and ObservabilityTable 23: Streaming vs Batch Ingestion Decision

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Architecture Layers

Understanding what work belongs in each layer is the foundation of a healthy medallion design. Each layer has a distinct contract: Bronze = truth, Silver = quality, Gold = value. Misplacing logic — especially pushing domain knowledge into Silver — is the single most common failure mode.

Layer	Example	Description
Bronze Layer	Raw JSON, CSV, Parquet ingested as-is	• Immutable landing zone for raw data • preserves original fidelity including errors, duplicates, and schema drift • append-only • never transformed
Silver Layer	Deduplicated, type-cast, null-checked records	• Conformed, validated, integrated data across sources • technical transformations only • no business logic requiring domain knowledge
Gold Layer	`fact_orders`, `dim_customer`, `report_revenue_by_region`	• Business-ready aggregations modeled for specific use cases • star schema, wide tables, or OBT patterns • optimized for BI and ML
Landing / Staging Zone	S3/ADLS raw dump folder, Event Hub checkpoints	• Optional pre-Bronze holding area for transient raw files • no Delta/Iceberg formatting required • cleared after ingestion to Bronze
Platinum Layer	Near-real-time ML-scored records, streaming aggregates	Optional fourth layer for highly curated ML-ready features, real-time scoring outputs, or board-level KPIs requiring strict SLAs.
Domain-Scoped Layers	`finance.silver`, `supply_chain.gold` catalogs	• Medallion Mesh pattern — each domain team owns its full Bronze→Gold pipeline • catalogs organized domain-first, not layer-first, for governance at scale

Table 2: Bronze Layer Ingestion Patterns

Bronze layer patterns determine how raw data is captured and preserved. The right pattern depends on source type, latency requirement, and downstream replay needs. Every pattern shares one invariant: no data transformation — only capture.

Pattern	Example	Description
Full Load	`spark.read.csv(path).write.format("delta").save(bronze_path)`	• Replace entire Bronze table on each load • simplest pattern • safe only for small, slowly changing source tables

The rest of this section is part of the full cheat sheet

Table 3: Bronze Layer Design Principles

These principles ensure Bronze tables remain reliable as the audit trail and replay source for all downstream layers. Violating them — especially allowing mutations — destroys trust in the entire pipeline.

Principle	Example	Description
Immutability	Append-only writes; no `UPDATE` or `DELETE` on Bronze	• Bronze must never be mutated • source data defects should be corrected in Silver, not Bronze, to preserve audit trail

The rest of this section is part of the full cheat sheet

Table 4: Silver Layer Transformation Standards

Silver transforms raw Bronze data into a reliable integration layer that any consumer across the organization can trust. The "new hire test" defines scope: transformations a day-one employee can perform without business knowledge belong in Silver; those requiring domain expertise belong in Gold.

Standard	Example	Description
Deduplication	`MERGE INTO silver.orders USING bronze.orders ON id WHEN MATCHED THEN UPDATE`	• Remove duplicate records introduced by retries or at-least-once delivery • use MERGE for idempotent upserts

The rest of this section is part of the full cheat sheet

Table 1: Core Architecture Layers

Layer	Example	Description
Bronze Layer	Raw JSON, CSV, Parquet ingested as-is	• Immutable landing zone for raw data • preserves original fidelity including errors, duplicates, and schema drift • append-only • never transformed
Silver Layer	Deduplicated, type-cast, null-checked records	• Conformed, validated, integrated data across sources • technical transformations only • no business logic requiring domain knowledge
Gold Layer	`fact_orders`, `dim_customer`, `report_revenue_by_region`	• Business-ready aggregations modeled for specific use cases • star schema, wide tables, or OBT patterns • optimized for BI and ML
Landing / Staging Zone	S3/ADLS raw dump folder, Event Hub checkpoints	• Optional pre-Bronze holding area for transient raw files • no Delta/Iceberg formatting required • cleared after ingestion to Bronze
Platinum Layer	Near-real-time ML-scored records, streaming aggregates	Optional fourth layer for highly curated ML-ready features, real-time scoring outputs, or board-level KPIs requiring strict SLAs.
Domain-Scoped Layers	`finance.silver`, `supply_chain.gold` catalogs	• Medallion Mesh pattern — each domain team owns its full Bronze→Gold pipeline • catalogs organized domain-first, not layer-first, for governance at scale

Table 2: Bronze Layer Ingestion Patterns

Pattern	Example	Description
Full Load	`spark.read.csv(path).write.format("delta").save(bronze_path)`	• Replace entire Bronze table on each load • simplest pattern • safe only for small, slowly changing source tables

Table 3: Bronze Layer Design Principles

Principle	Example	Description
Immutability	Append-only writes; no `UPDATE` or `DELETE` on Bronze	• Bronze must never be mutated • source data defects should be corrected in Silver, not Bronze, to preserve audit trail

Table 4: Silver Layer Transformation Standards

Standard	Example	Description
Deduplication	`MERGE INTO silver.orders USING bronze.orders ON id WHEN MATCHED THEN UPDATE`	• Remove duplicate records introduced by retries or at-least-once delivery • use MERGE for idempotent upserts