Databricks · One-Pager · Architecture

Lakehouse & Medallion

One platform that fuses a data lake's cheap, open storage with a warehouse's reliability and governance — and a Bronze→Silver→Gold pattern for refining data inside it.

GA · Reference architecture Delta Lake + Unity Catalog Verified Jun 2026 docs

In one line

A lakehouse is a single data-management system that combines the benefits of data lakes and data warehouses; medallion is the layered data-design pattern (Bronze → Silver → Gold) for progressively improving data quality within it.

Why it matters

It removes the classic two-system split — a data lake for ML plus a separate warehouse for BI — so one open copy of the data, under one governance model, serves analytics, BI, and AI without redundant infrastructure or brittle copies. Medallion gives teams a shared vocabulary for where in the pipeline a table sits and how trustworthy it is.

Core concepts

Lakehouse

Data-management system combining data-lake cost/flexibility with data-warehouse query optimization and reliability — no separate systems.

Delta Lake

The open storage layer that makes it work: ACID transactions and schema enforcement on cheap object storage.

Unity Catalog

Unified, fine-grained governance for data and AI — access control and lineage tracking back to the source of truth.

Bronze (raw)

Raw, unvalidated data in its original form — the single source of truth that enables reprocessing and auditing.

Silver (validated)

Cleansed, deduplicated, conformed data with schema enforcement and types applied — the queryable, joined enterprise view.

Gold (enriched)

Refined, aggregated, business-level tables optimized for analytics, reporting, and dashboards.

How it works

Click a layer · data quality improves left → right

0 · Sources 1 · Bronze 2 · Silver 3 · Gold

Sources — operational systems, files in cloud object storage, databases, and streams land in their native format.

Minimal example

SQL · one schema per layer in a Unity Catalog catalog-- Silver: clean + dedup the raw Bronze table into a validated table
CREATE OR REFRESH STREAMING TABLE ops.silver.orders AS
SELECT DISTINCT
  CAST(order_id AS BIGINT) AS order_id,
  CAST(amount   AS DECIMAL(12,2)) AS amount,
  customer_id, order_ts
FROM STREAM(ops.bronze.orders)   -- incremental read of the raw layer
WHERE order_id IS NOT NULL;

Gotchas & limits

Medallion is a recommended pattern, not a requirement and not enforced — Databricks won't stop you from skipping a layer or adding a fourth.
"Bronze/Silver/Gold" is convention, not a product feature: the platform doesn't track which layer a table is; you organize it (often one schema per layer).
Don't over-build: trivial pipelines don't need three full layers, and copying data three times has storage and latency cost — match layer count to need.
The lakehouse depends on Delta Lake for ACID + schema enforcement and Unity Catalog for governance; plain Parquet on a lake gives you neither.

Interview soundbites

Say this out loud

"A lakehouse is one system that combines a data lake and a data warehouse — Delta Lake gives ACID and schema enforcement, Unity Catalog gives governance, so one open copy of the data serves BI, ML, and AI."
"Medallion is a multi-hop pattern — Bronze is raw and immutable, Silver is cleansed and conformed, Gold is business-level aggregates — and data quality improves as it flows Bronze to Silver to Gold."
"It's a recommended best practice, not a hard rule — the layer names are convention; I'd map them to Unity Catalog schemas and size the number of layers to the use case."

References

Verified against docs.databricks.com · June 2026