TIL: Medallion Architecture is a layered way to organize data pipelines so data quality improves step by step.
What is Medallion Architecture?
It is a pattern where data moves through 3 layers:
- Bronze: raw ingested data
- Silver: cleaned and validated data
- Gold: business-ready aggregated data
Simple idea: do not jump directly from raw to dashboards. First keep raw, then clean, then model for analytics.
Bronze layer (raw)
Bronze stores source data as-is with minimal transformation.
Typical characteristics:
- append-only ingestion
- full raw records
- useful for reprocessing and audit/debug
Think of this as your “ground truth landing zone.”
Small example (Bronze):
order_id=101, user_id=42, amount="1999", created_at="2026-06-16T10:11:12Z", source="mobile-app"order_id=101, user_id=42, amount="1999", created_at="2026-06-16T10:11:12Z", source="mobile-app" (duplicate)
Silver layer (refined)
Silver is where data quality improvements happen.
Typical operations:
- schema enforcement and normalization
- deduplication
- null handling and basic validations
- join/reference fixes
At this stage, data is trustworthy enough for most internal analytics and feature prep.
Small example (Silver):
order_id=101, user_id=42, amount=1999.00, created_date=2026-06-16, channel="mobile"
- duplicate removed
- amount converted to numeric type
- timestamp normalized to a standard date/time format
Gold layer (curated/business)
Gold is optimized for consumption by BI tools, reports, and product metrics.
Typical operations:
- aggregations (daily, weekly, monthly)
- dimensional/star style modeling
- KPI-ready tables (for example: revenue by region, retention cohorts)
Gold should be stable, easy to query, and aligned with business definitions.
Small example (Gold):
date=2026-06-16, channel="mobile", total_orders=12450, total_revenue=24500000
This is now directly usable in dashboards/KPI reporting.
Why this model works well
- easier debugging (you can inspect each stage)
- safer transformations (quality gates between layers)
- reusability (many gold datasets can reuse same silver data)
- better governance and lineage
Quick mental model
Raw events/files -> Bronze -> Silver -> Gold -> BI/ML/Apps
That is the core of Bronze/Silver/Gold in data engineering.