Databricks Clean Rooms: Privacy-Preserving Data Collaboration

Databricks Clean Rooms is a privacy-preserving feature that lets multiple organizations collaborate on shared data without ever exchanging raw rows. Built on top of Unity Catalog and Delta Sharing, it ensures that each participant's tables only return the results of approved aggregate queries.

Adoption is growing in industries where direct exchange of personal data is not allowed: ad reach measurement, cross-bank fraud detection, and healthcare epidemiology. This article walks through how Clean Rooms work, how to design them, and how they compare with Snowflake.

Clean Rooms Overview

A clean room is the data privacy concept of a 'safe compute environment'. Two or more organizations bring their own data and can join, aggregate, and analyze it without ever accessing each other's raw records.

Databricks Clean Rooms is built on the following technology stack:

• Unity Catalog: foundation for table access control (GRANT / DENY) and audit logs
• Delta Sharing: the cross-organization data sharing protocol that mounts Delta tables into the recipient's Databricks workspace
• Secure Computation: the layer that restricts computation on shared data so only approved aggregates can run

Architecture and Flow

The Clean Rooms flow consists of the following steps:

[Organization A]              [Clean Room]              [Organization B]
      |                          |                          |
      |-- Publish table -------->|                          |
      |   (customer_id, orders)  |                          |
      |                          |<-------- Publish table --|
      |                          |   (customer_id, ad views)|
      |                          |                          |
      |                          |-- Join + aggregate ----->|
      |                          |   (JOIN ON hashed_id)    |
      |                          |                          |
      |<-- Aggregate result only-|                          |
      |   (reach, CVR, ...)      |                          |
      |                          |------- Aggregate only -->|
      |                          |                          |
  * Raw data never leaves the owner

Collaborator Setup

Participants (collaborators) join a Clean Room via the following steps:

• Clean Room creator: the organization that creates the Clean Room and invites participants. Manages the clean room object in Unity Catalog.
• Adding collaborators: invite the other organization by pointing at their Databricks metastore via Delta Sharing.
• Publishing tables: each participant registers their own tables in the Clean Room and defines the shared columns and aggregation rules.

Table Sharing Rules

Apply the following rules to tables published to a Clean Room to prevent data leakage:

• Restrict shared columns: share only the columns required for the analysis, not the whole table. Hash personally identifiable columns (name, email, etc.) or exclude them entirely.
• Aggregation-only constraint: restrict queries to return only aggregate results (COUNT / SUM / AVG), never row-level data.
• Minimum count threshold (k-anonymity): suppress results whose group size is below k. Reduces the risk of identifying individuals from small groups.
• Limit join keys: restrict which columns are allowed in JOIN conditions. Typically only a hashed customer ID is permitted.

Differential Privacy

Differential privacy adds mathematically calibrated noise to aggregate query results, making it statistically impossible to tell whether any specific record is in the dataset.

Differential privacy in Databricks Clean Rooms works as follows:

• Epsilon parameter: controls the level of privacy. Smaller epsilon means more noise and stronger privacy, but lower analytical accuracy.
• Privacy budget: caps the number of queries against the same dataset, preventing inference attacks via repeated queries.
• Noise mechanisms: Laplace or Gaussian noise is added to aggregate values, statistically masking the influence of any individual record.

Key Use Cases

Comparison with Snowflake Clean Rooms

Key Exam Points

• "Two companies want to share data but cannot expose raw rows" → Clean Rooms
• "What is the foundation tech of Clean Rooms?" → Delta Sharing + Unity Catalog
• "Difference between Delta Sharing and Clean Rooms?" → Delta Sharing exposes raw data; Clean Rooms expose only aggregates
• "What techniques inside Clean Rooms prevent re-identification?" → differential privacy / k-anonymity-style thresholds

Check with a Sample Question

Frequently Asked Questions