The Evolution of Booking.com's AI Journey: A Deep Dive into the Unpolished Story
Jabez Eliezer Manuel's presentation at QCon London 2026 offers a captivating glimpse into Booking.com's remarkable AI evolution. It's a story of relentless innovation, overcoming challenges, and a unified command center that drove their success.
The Journey Begins: A Leap Back in Time
Manuel takes us on a trip down memory lane, back to 2005. The Motorola Razr V3 was a popular cell phone, Web 2.0 was taking its first steps, and Booking.com was just nine years old. This era marked the beginning of their A/B testing experiments, a crucial phase in their data-driven DNA development. With over 1000 experiments in parallel and 150,000 total experiments, the goal wasn't perfection but rapid learning.
Data Management: From Simplicity to Scale
Booking.com's initial tech stack, built on Perl libraries and MySQL, served them well for a time. But as their data grew, so did the need for a more robust solution. They embraced Apache Hadoop for distributed storage and processing, scaling to two on-premise clusters with 60,000 cores and 200 PB of hard disk space by 2011. However, Hadoop's limitations soon became apparent.
Noisy neighbors, lack of GPU support, and capacity issues plagued their machine learning pipeline. The decision to sunset Hadoop in 2018 was a strategic move, but the migration process took seven years. A meticulous five-phase strategy, including ecosystem mapping, scope reduction, and interleaving A/B testing, ensured a smooth transition.
Machine Learning Engineering: A Transformative Journey
Booking.com's machine learning stack evolved significantly. From Perl libraries and MySQL to Apache Oozie with Python, Apache Spark with MLlib, H2O.ai, deep learning, and GenAI, their platform transformed. 2015 proved pivotal, solving real-time predictions and feature engineering.
Today, their inference platform handles over 480 machine learning models, 400 billion predictions daily, and boasts a latency of less than 20 milliseconds. This level of sophistication is a testament to their commitment to continuous improvement.
Domain Intelligence: Personalization at Scale
Manuel highlights four domain-specific platforms: GenAI for trip planning and smart filters, Content Intelligence for image and review analysis, and Recommendations for personalized content. But the most complex challenge lies in Ranking.
Booking.com's ranking formula, initially a simple function, evolved to include factors like cancellations, distance, and availability. However, replacing it with machine learning proved difficult due to infrastructure limitations. Interleaving A/B testing experiments allowed for more variants with less traffic, leading to a more refined approach.
The Unified Command Center: A Key to Success
The presentation emphasizes the importance of a unified command center. This centralized hub, accessible to all stakeholders, facilitated collaboration and informed decision-making. It enabled Booking.com to orchestrate their complex AI infrastructure effectively.
Looking Ahead: The Future of Booking.com's AI
Manuel concludes by highlighting the unification of domain-specific platforms under a single orchestration layer. This integration paves the way for even more sophisticated AI applications, further enhancing the customer experience and driving Booking.com's continued success in the ever-evolving travel industry.
Reflection and Takeaway
Booking.com's journey is a testament to the power of continuous learning, adaptation, and a unified vision. Their story serves as an inspiration for other organizations embarking on their own AI journeys, emphasizing the importance of a holistic approach and a unified command center in navigating the complexities of AI implementation.
