TUI Moves Faster and Builds Better ML Models Using MLOps from Data Reply on AWS

Established in 1923, the TUI Group owns—in whole or part—400 hotels, 16 cruise ships, and a digital platform for more than 160,000 tours, activities and experiences, and tour operators. It also owns 1,200 travel agencies and online portals, and five airlines with around 130 aircraft. Operating on such a large scale, the company embraced technology and had multiple teams working in related areas across the organization. The company needed to address that when it decided to rethink its approach to machine learning.

TUI has several different data science teams: one in the marketing area, focused on customer, domain, and digital projects; a commercial data science team responsible for things like pricing; and an airlines data science team. It also has a common data science team that acts as a central resource. And it has other teams specific to particular business units, such as those handling excursions. “It’s different teams often working in different areas,” says Dominic Rehn, head of digital data science at TUI. “We use machine learning in a lot of ways. Most prevalently, it’s probably around the standard package business. That was the focus in the short term due to the large scale. Now we can focus on other parts of the business.”

TUI needed to create stronger personalized experiences, in near-real time and at scale, to convert website visitors to customers. “We had to make sure that we could develop models for the different product owners and get those models to them quickly,” says Rehn. “We needed to be focused on the actual business issues rather than on infrastructure.” Speed to market is critical for TUI’s agility and competitive advantage.

However, TUI data scientists did not have an efficient way of working. Different teams in different geographies used a variety of open-source tools. Without a standard development environment, it was difficult and time consuming to deploy models in production, diverting data scientists’ focus from model creation and optimization.

To address these challenges, TUI wanted to adopt an MLOps approach to ML model development and deployment that follows the same philosophy of continuous development seen in DevOps.

TUI evaluated several companies and chose Data Reply to help it on this journey. Deciding factors were Data Reply’s expertise with Amazon Web Services (AWS) and its deep knowledge of MLOps, including its MLOps Accelerator. “We were already using AWS, so we needed someone who understood that and, at the same time, had real experience and competence in ML operations,” says Rehn.

The project started well because TUI was able to clearly express what it wanted to achieve. “TUI wanted to operationalize their machine learning,” says Aditya Jain, machine learning operations consultant at Data Reply. “They wanted more machine learning models in production. They wanted a stable framework in order to build these models, version them, deploy them in a sustainable way, and monitor them. They had clear goals and MLOps was the right approach to achieve those goals.”

Data Reply and TUI worked together in a three-phase project that included an assessment, implementation, and review to create an MLOps solution customized to TUI’s requirements. By working in collaboration with TUI’s data scientists and engineers, Data Reply was able to understand how the company worked and build a truly bespoke solution on AWS using existing tools and processes where appropriate, rather than replacing everything and starting from scratch. “It was a very effective use of our time,” says Rehn. “Data Reply focused on reviewing relevant processes, not trying to reinvent everything. The options they presented were relevant and useful.”

The solution was built using Amazon SageMaker, a fully managed service that brings together a broad set of tools to enable high-performance, low-cost machine learning for any use case. Data Reply also implemented the Amazon SageMaker Feature Store—a fully managed, purpose-built repository to store, share, and manage features for machine learning models. “The feature store allowed us to try things out and make sure that we understand how things work and if they’re a benefit to us without incurring big costs,” says Rehn. “We could quickly see if we wanted to continue with it.”

TUI uses other AWS services to support its data scientists and MLOps, including Amazon EC2 Auto Scaling, which helps maintain application availability and lets you automatically add or remove EC2 instances, and Amazon Simple Storage Service (Amazon S3), an object storage service offering industry-leading scalability, data availability, security, and performance. Amazon CloudWatch—a service that monitors applications, responds to performance changes, optimizes resource use, and provides insights into operational health—also helps TUI ensure that things are working as they should.

The three-step process of assessment, implementation, and review took only about 6 months from start to finish and delivered multiple benefits. TUI created an MLOps engineer role to help bring the work done by the data scientists to production, rather than having its data scientists handle everything from development to production. This enables data scientists to focus more on business value. It also helped to improve data culture within TUI—the business is more aware about what it takes to build products rather than simply do projects.

Simplifying architecture and workflow has also helped to reduce the time required to onboard new data scientists by 75 percent—to just 2 weeks from 2 months. The process has also reduced model development and deployment time by about 66 percent, leading to 10 new models going live in just the initial 6 months.

Since implementing the MLOps solution, TUI has been able to work on hundreds of models. Those models are making a difference to TUI’s bottom line. “In our use cases that we’ve delivered via the platform so far, we’ve put out models that are in the region of around €6 million to €7 million margin a year in benefit for us,” says Rehn. “We’re looking at plans now on how we can roll that out even further to around €20 million a year.”

Although the benefits are impressive, the process of working with Data Reply was just as impressive. “A lot of consultancies just want to come in and rebuild your whole company,” says Rehn. “What we got from Data Reply wasn’t just an architecture diagram. They looked at how we worked, looked at the roles in the process, and suggested ways to improve and helped us understand what we were missing. Data Reply worked side by side with us to build an MLOps approach that’s right for us and we’ll be expanding throughout the company.”