For decades, the computing and travel sectors have been inextricably linked. To put things in context, American Airlines secured an agreement with IBM in 1957, and the two IBM 7090 mainframes were used to construct the first computer reservation system in 1960.

That same booking system eventually evolved into Sabre, one of the three main global distribution systems (GDS), which trades flights from a variety of airlines and distributes them to travel agents and consumers.

However, when it comes to providing 21st-century customers, that tradition poses a dilemma.

"Data formats in the travel business have been around for decades," says Andrew Gasparovic, chief architect of Sabre Labs, the GDS's technology section, which was created in 1996.

They were created at a time when there wasn't much thought put into what data may be used. "What we term the 'offer and order' model didn't work with them," he told The Register.

In e-commerce, bringing data about customer offers and the reservations they eventually make is typical, and it helps vendors predict what customers will buy next.

However, according to Gasparovic, bringing that data together across more than 12 billion purchasing requests and 1 billion travelers each year was not an easy effort and one that was assigned to cooperation with Google in 2020.

Sabre is preparing to move its IT infrastructure to Google Cloud. It's also implementing operational data technologies like managed systems Spanner, a distributed database that powers GoogleAds, and BigTable, a wide-column, key-value NoSQL database.

It uses BigQuery, Google's distributed data warehouse, for analytics.

Amadeus GDS, founded in 1987 by Air France, Lufthansa, Iberia, and SAS airlines as a European alternative to Sabre, and Travelport, located in the United Kingdom, are two other GDSs (which include the Apollo, Worldspan, and Galileo GDS.) All of these networks started processing airline and travel agent ticketing, but now they also deal with travel websites, car rental companies, and hotels.

Sabre's first travel distribution network existed before the internet was invented. A passenger's booking information, seat selection, tickets, special requests, and other vital information regarding their trip are all stored in an airline's reservation database. On behalf of carrier customers, Sabre generally handles thousands of reservation modifications per second. The reservation database of an airline must be supplied from a variety of availability zones.

Meanwhile, employing smartphone apps, third-party travel websites, and airline call centers, the flight shopping system generates millions of itineraries each second on behalf of travelers. It uses Bigtable to manage 10 exabytes of data. Building a data solution to predict what customers could buy next is built on top of Teradata, Oracle, and IBM's existing data warehousing infrastructure.

"In terms of data warehousing technologies and operational data storage, we have pretty much everything you'd expect. We're considering how to get a feed into BigQuery for all of those existing systems "According to Gasparovic.

"It was difficult to understand a traveler when they transitioned from looking for tickets to booking flights. To use that data to figure out what those travelers are interested in, what their tastes are, and what kinds of product bundles they typically purchase. Because the data was stored in so many different systems and so many various formats, this was a difficult task "he continues.

Sabre Labs keeps track of what was offered to the passenger, providing an offer ID that is fed into BigQuery with the customer's order. "It's not just a matter of putting it in the same spot. It's about updating and modernizing the data model itself so that it can identify those objects using a unique key "he declares

The initial stage was to combine everything into one system, start correlating it, and comprehend it as a whole. The next step was to create data-driven machine learning models. "This is a huge step forward in terms of what we can offer our clients," Gasparovic says.

Machine learning, on the other hand, comes with a cautionary note. It's possible to tinker with machine learning for hours on end without ever solving a genuine problem. According to him, projects must begin with real-world challenges to solve.

Sabre Labs uses a reinforcement learning technique to compress the training portion of a machine learning project, allowing them to quickly try something out and assess what advantage it gives. "Think of it as a fancier version of A-B testing with a lot of different elements being tried at the same time," Gasparovic explains.

"The benefit of this is that we don't have to spend a lot of time upfront training models and tinkering with parameters to achieve accuracy. When we've established that it's useful, we can return to more traditional machine learning methods, such as building models and performing supervised learning, in which you train a model on data, then put it into production, and use its outputs to make predictions, but starting with that experimentation has been extremely beneficial in determining where we want to spend our time "he declares

It has already offered machine learning models to its partner company to aid in the creation of better travel offerings for customers.

"It's something that customers expect when we can put that on a website and see the impact, but it's been successful for hotels and airlines. The fact that we were able to make this happen for them is still unique in the industry "According to Gasparovic.

The Google Cloud solution is built on top of Sabre's current data warehouses, but the objective is to consolidate data onto one system in the long run. For "some time," Sabre will likely be operating both worlds, with the new world layered on top of the old, he says.

Meanwhile, Gasparovic expects the current setup to produce more victories. "We've merely scratched the surface of what machine learning and experimentation can do," adds Gasparovic.