Booking.com's 20-Year Journey Building Its Data and ML Platform
Booking.com runs a machine learning platform that produces 400 billion predictions per day. Jabez Eliezer Manuel walks through the 20-year evolution of the company's data and ML platforms, from database scaling to its newest agentic AI foundation.
From MySQL and Hadoop to a cloud architecture
On the data management side, Jabez explains how the Booking.com team scaled its MySQL topologies. They also sunset Hadoop and moved to a cloud architecture. The technical details behind that decision are covered in the presentation itself.
A machine learning platform at 400 billion predictions
On the ML side, the team built high-throughput model serving and a feature store. Both operate at a scale of 400 billion predictions per day. A feature store is a place to store and serve the data features that models use, so many models can share them.
Jabez also shares insights on personalization and interleaved experimentation. This experimentation method typically blends the results of two models into a single view, then compares which one users prefer. At the end of the session, he introduces the new agentic AI foundation Booking.com is building.
If you manage a digital product or a data team, the story is worth a look because it shows a platform this large was built step by step over two decades. The move from Hadoop to the cloud is also a real example that older data architectures can, and sometimes must, be replaced as needs change.