Problem
Warner Bros. Discovery (WBD) kept customer data for several brands in separate silos, so marketers could not build one audience across them. Privacy rules (GDPR, CCPA) required tight control of data use, and an off-the-shelf CDP did not fit the scale WBD needed. Marketers also want audience sizes while they edit rules, but on warehouse tables a simple audience took about one minute and a complex one more than five.
Example from the article: a marketer builds an audience of people who engaged with Superman and Wonder Woman in the last 12 months to promote the Supergirl release, and activates it to Meta paid media.
Solution in brief
WBD built its own CDP, Mosaic, on Databricks with three planes. The UI stores audience rules as metadata, a rule compiler turns them into Spark SQL that Databricks runs, and a separate service delivers results to partners. Identity resolution and consent filtering run inside the lakehouse. A sample-based estimate, corrected by XGBoost, gives audience sizes in under 30 seconds.
Questions this case answers
- How to compile user-defined selection rules into SQL on a lakehouse so that segment size estimates return in seconds instead of minutes
- How to apply consent and opt-out filters inside the data platform before results reach downstream systems and how to resolve identities across user, household and device records
- How to build a customer data platform (CDP) on a lakehouse instead of buying an off-the-shelf one
- How to compile drag-and-drop audience rules into SQL that runs on a lakehouse
Tools and methods
- Three-plane architecture — metadata, execution, delivery
- Rule compiler (Nexor) — drag-and-drop rule tree to DNF to Spark SQL
- Delta Lake, Spark Structured Streaming, Spark SQL — ingestion and audience evaluation
- Identity tables in Delta — identity graph as a relationship table plus an identity table
- Sample, scale, calibrate — audience size estimate with XGBoost on MLflow
- Syndra (Akka actors) — async delivery to partner destinations
Google or a one-time email code. No password.