In the ever-evolving landscape of technology, it's fascinating to trace the journey of a company like Booking.com, which has seamlessly woven AI into its very fabric. The story of Booking.com's AI evolution, as presented by Jabez Eliezer Manuel at QCon London 2026, is a testament to the power of adaptability and innovation. What makes this narrative particularly intriguing is the unpolished, behind-the-scenes look at the challenges and triumphs that shaped their data-driven DNA. As I delved into the presentation, a few key insights stood out, each with its own unique flavor of commentary and analysis.
The Data Management Odyssey
Booking.com's data management story is a classic tale of growth and evolution. Starting with a single master database in 2005, they scaled to approximately 6800 database instances by 2020. What's truly remarkable is their unique setup, which includes smaller databases (with a 2TB limit) fitting into Non-Volatile Memory Express (NVMe) solid-state drives. This allowed for lightning-fast point queries, a mere 350 microseconds. However, as their data grew, they hit a wall, prompting the addition of Apache Hadoop for distributed storage and processing. The migration process, spanning seven years, was a complex dance, with five distinct phases, including mapping the ecosystem, analyzing usage, and applying the Google Search PageRank algorithm. The key takeaway here is the importance of a unified command center, which guided their successful transition away from Hadoop.
Machine Learning Engineering: A Journey from Perl to GenAI
Booking.com's machine learning stack has undergone a remarkable evolution, starting with Perl libraries and MySQL in 2005 and culminating in agentic systems by 2025. The pivotal year of 2015 marked a turning point, solving real-time predictions and feature engineering challenges. Today, their inference platform boasts over 480 machine learning models, making over 400 billion predictions daily with a latency of less than 20 milliseconds. This is a testament to the power of iterative improvements and the importance of addressing core problems. What's particularly fascinating is the interleaving technique they employed in A/B testing, allowing for more variants with less traffic, a strategy that could be a game-changer for many.
Domain Intelligence: From Ranking to Personalization
Booking.com's domain-specific machine learning platforms showcase their commitment to innovation. From GenAI for trip planning and smart filters to Content Intelligence for detailed hotel content and Recommendations for personalized content, each platform serves a unique purpose. However, the ranking platform stands out as a complex three-way optimization challenge, balancing choice and value, exposure and growth, and efficiency and revenue. The evolution of their ranking formula, from a simple function to a sophisticated machine learning model, highlights the challenges of replacing a seemingly 'undefeatable' formula due to infrastructure limitations. This journey underscores the importance of iterative improvements and the need to adapt to changing circumstances.
The Unpolished Story: A Reflection
What makes Booking.com's story so compelling is the unpolished, behind-the-scenes look at their journey. It's a reminder that even the most successful companies face challenges and make mistakes. The key lies in learning from these experiences and adapting to change. Booking.com's ability to evolve its data management, machine learning engineering, and domain intelligence platforms showcases the power of innovation and the importance of a unified command center. As we reflect on their journey, it's clear that the future of AI in the travel industry is bright, and Booking.com is at the forefront of this exciting evolution. Personally, I find it inspiring to see how a company can seamlessly weave AI into its DNA, and Booking.com's story is a testament to the power of adaptability and innovation.
