/fabric-medallion — Medallion Architecture on Fabric

Medallion is the standard Fabric OneLake data pattern: raw at bronze, cleansed at silver, business-ready at gold. Each layer has explicit quality gates and promotion rules. This skill produces the zone definition and the contracts between them.

When to invoke

• Kick-off of a Fabric project where raw source data needs a durable layout
• Atlas is designing a lakehouse before ingestion begins
• Bob is architecting analytics that will consume multiple domains
• A downstream consumer (Power BI, notebook, ML model) needs a stable business-ready table

Steps

1. Define the bronze zone (raw). Land source data as-is — original schema, original casing, original nulls. Delta Lake format. Partition by ingestion date. Table naming: bronze_<source>_<entity>. No transformation beyond append.
2. Define the silver zone (cleansed). Deduplicate, enforce types, resolve reference codes, apply late-arriving-data policy (SCD-1 or SCD-2 per entity). Table naming: silver_<domain>_<entity>. Silver is queryable by analysts but not yet aggregated for business logic.
3. Define the gold zone (business-ready). Star-schema facts and dimensions, aggregations, business-owned KPI logic. Table naming: gold_<subject_area>_<fact_or_dim>. Gold is what Power BI semantic models point at.
4. Promotion policy. Explicitly state: what triggers bronze → silver (schedule + data-arrival), silver → gold (source freshness + validation), and rollback rules. Never let gold reference silver mid-pipeline without a lock.
5. Quality gates. Between each layer: schema validation, null-rate thresholds, referential integrity checks, row-count sanity, freshness SLA. Fabric Data Quality (or Great Expectations if custom) — verify current tool at author time. Failed gate blocks promotion.
6. Data contracts. For each gold table, publish a contract: column names, types, nullability, freshness SLA, owner, upstream lineage. Contract lives in Git alongside the pipeline code.

Deliverable format

• medallion.md — zone definitions with naming conventions and partitioning
• contracts/<table>.yaml — one per gold table, with schema and SLA
• Fabric pipeline YAML or notebook that enforces promotion rules
• Dashboard link (Fabric Monitoring Hub) showing gate pass/fail per run

Non-negotiables

• Bronze is immutable — never mutate landed data, only append
• Gold never reads from bronze directly — always through silver
• Every gold table has a data contract in Git before it is consumed downstream
• Quality gate failure blocks promotion — no manual override without a decision log entry

References

• Fabric — medallion lakehouse architecture
• Fabric — Delta Lake table format
• Fabric — Data Quality (preview at author time — verify)
• Superteam ADR — MS Fabric skills pack