Behind Booking.com's AI Evolution: The Unpolished Story

The presentation explores the evolution of Booking.com's AI architecture from its early days to present times.

The presentation is structured into three layers:

• Data Management Layer:
  • Evolution of database management, highlighting the transition from early technologies such as Perl scripts and MySQL to sophisticated cloud solutions.
  • Challenges in data cataloging and ownership, leading to discovery and management nightmares.
  • The significant migration journey from Hadoop to cloud infrastructure.
• ML Engineering Layer:
  • Model serving advancements that support over 480 models, enabling 400 billion predictions daily with sub-20ms latency.
  • The persistent challenges in feature engineering, highlighting attempts to standardize and centralize feature computations and storage.
• Domain Intelligence Layer:
  • Overview of specialized ML platforms developed for various purposes such as Ranking, Recommendations, and Content Intelligence.
  • Innovations like the introduction of generative AI platforms for scalable, intelligent product development.

Key lessons shared include the importance of experimentation culture, data centralization, and strategic decision-making in AI evolution. Additionally, future plans emphasize further unification of platforms and developing agentic systems to enhance the AI capabilities of Booking.com.

Jabez concludes with recommendations on asset cataloging, governance, and the significance of taking a unified approach to infrastructure management.

This is the end of the AI-generated content.

It’s easy to look at a mature AI platform and imagine a grand blueprint. Ours began with none. What started as a few data scientists hacking on Perl scripts and Mysql queries has grown into an AI platform that impacts millions of travel decisions every day. This is the story of that transformation from an engineering perspective.

This talk traces Booking.com's AI architecture evolution over the last two decades—from deterministic formulas and cron jobs to specialized ML platforms for Ranking, Recommendations, Content Intelligence and GenAI. We'll focus on the backend and infrastructure choices that made that journey possible: a unique MySQL setup that scales without caching, the painful seven-year migration from Hadoop to cloud, real-time ML inference at scale, and our ongoing struggle with feature engineering.
No data science deep dives—just real engineering trade-offs, missteps, and hard-won lessons.

You'll learn:

• How experimentation culture became our foundation for data-driven decisions
• Why we run MySQL at scale without a single cache layer
• The cost of not cataloging your data: discovery and ownership nightmares
• Why knowing that a problem exists doesn't mean you can solve it quickly (our Hadoop story)
• Practical patterns for migrating petabyte-scale data infrastructure
• Architecture patterns for serving billion ML predictions daily with sub-20ms latency
• The feature engineering challenge: three attempts, still no silver bullet
• A peek inside the Agent Catalog that powers our GenAI use cases
• When ML ranking models couldn't beat a hand-coded formula