Booking.com’s AI Evolution: Insights from Jabez Eliezer Manuel at QCon London 2026
In a gripping presentation at QCon London 2026, Jabez Eliezer Manuel, Senior Principal Engineer at Booking.com, unveiled “Behind Booking.com’s AI Evolution: The Unpolished Story.” This talk offered a deep dive into the 20-year journey of Booking.com as it has innovated and navigated the challenges of AI integration within its vast operational landscape.
A Journey Through Time: Booking.com’s Technological Evolution
Kicking off the session, Manuel took the audience on a nostalgic journey back to 2005. At that time, the Motorola Razr V3 was the epitome of mobile technology, and Web 2.0 was just beginning to reshape our online experiences. Here, Booking.com was a youthful nine years old, laying the groundwork for its future innovations.
It was in February 2005 that Booking.com embraced its first foray into A/B testing, conducting over 1,000 experiments concurrently. With a staggering total of 150,000 experiments to date, the company realized a success rate of under 25%. However, Manuel emphasized that the true objective wasn’t always about being correct; rather, it was about rapid learning, which forged Booking.com’s “Data-Driven DNA.”
Data Management: Laying the Foundation
The backbone of Booking.com’s technological prowess lies in its robust data management architecture, which has evolved dramatically over the years. Initially grounded in Perl libraries and MySQL, the company’s setup featured asynchronous replication and commercial backing. Back in 2005, there was a single master database which, by 2020, transformed into approximately 6,800 instances.
Interestingly, their approach defied convention—eschewing specialized hardware, stored procedures, and even cache layers. Manuel described their “secret sauce” as the use of smaller databases, constrained to a 2TB limit, deployed in Non-Volatile Memory Express (NVMe) solid-state drives. This setup optimized point queries to an impressive average of less than 350 microseconds.
More Read
However, as the enormity of data burgeoned, Booking.com encountered new challenges. To scale effectively, they integrated Apache Hadoop, setting up two on-premise clusters in 2011 that boasted around 60,000 cores and 200 petabytes of storage. Yet, as Manuel pointed out, over time, issues such as noisy neighbors and capacity limitations led to congested clusters, prompting the decision to phase out Hadoop by 2018. This migration process, which spanned seven years, involved meticulously structured phases encompassing ecosystem mapping, usage analysis, algorithm application, wave migrations, and ultimately, Hadoop phase-out.
The Rise of Machine Learning Engineering
Booking.com’s journey through machine learning began with its humble Perl and MySQL roots, ultimately evolving into complex agentic systems by 2025. The progression involved pivotal transitions through various frameworks such as Apache Oozie, Python, and Apache Spark with MLlib. A significant milestone occurred in 2015 when the company successfully addressed core challenges related to real-time predictions and feature engineering to achieve online inference at scale.
Today, their machine learning inference platform is a marvel in itself, housing over 480 models, generating 400 billion predictions every day, all while maintaining latency below 20 milliseconds. This monumental advancement places them at the forefront of AI-driven customer experience optimization.
Domain Intelligence: Tailoring AI to Specific Needs
Manuel’s presentation delved into four specific machine learning platforms developed for distinct use cases within Booking.com. The first three include:
-
GenAI: Focused on trip planning, smart filters, and summarizing reviews, this platform enhances the user journey.
-
Content Intelligence: Serving as a data-rich machine learning content hub, it analyzes images and reviews while generating text relevant to hotel offerings.
-
Recommendations: This platform delivers personalized content to customers, enhancing the relevance of search results and bookings.
The fourth platform, Ranking, presented unique complexities with its three-way optimization challenges, navigating between choice, exposure, and efficiency. Manuel shared insights into how the initial ranking formula from 2005 relied on basic parameters, incorporating randomness. However, attempts to improve this model with machine learning ran into limitations due to infrastructure constraints.
To perfect their ranking strategy, the company innovated by interleaving experiments, merging 50% of different experimental sets into a single test. This method allowed for a broader range of variations without overwhelming traffic and ultimately validated their findings through targeted A/B testing.
As Booking.com continues to evolve, its rich history and ongoing innovations serve as a testament to the power of data-driven decision-making and the relentless pursuit of AI excellence. Through strategic adaptation and intelligent frameworks, the company has successfully navigated the complexities of integrating AI into its business model and continues to set benchmarks in the industry.
Inspired by: Source
