Travel tech sheds legacy baggage, heads to the cloud with Google

Sabre deal with Google creates parallel data architecture to support ML-based ecommerce partnerships with airlines and hotel chains

Feature The computing and travel industries have traveled hand in hand for decades. For perspective, American Airlines signed a deal with IBM in 1957 which developed the first computer reservation system in 1960, based on two IBM 7090 mainframes.

Over time, that same booking system morphed into one of the three largest global distribution systems (GDS), called Sabre, which trades flights from many airlines, supplying them to travel agents and consumers.

But with that heritage comes a problem when it comes to serving 21st-century consumers.

"The data formats that existed in the travel industry have been around for a very long time; for decades," says Andrew Gasparovic, chief architect at Sabre Labs, the GDS's technology division, founded in 1996.

"They were designed at a point in time when there wasn't a lot of thought given to what you could do with data. They weren't amenable to what we call the 'offer and order' concept," he told The Register.

Bringing data about customer offers and the reservations they finally make is common in ecommerce and helps sellers anticipate what customers are likely to buy next.

But, according to Gasparovic, getting that data together across more than 12 billion shopping requests and 1 billion travelers every year was not a trivial task and one given to a partnership with Google formed in 2020.

Sabre is in the process of migrating its IT infrastructure to Google Cloud. It is also adopting operational data tools including managed systems Spanner, the distributed database that backs GoogleAds and BigTable, the wide-column and key-value NoSQL database.

In analytics, it is using BigQuery, Google's distributed data warehouse.

Other GDSes include Amadeus GDS – created in 1987 by Air France, Lufthansa, Iberia and SAS airlines as a Europe-based alternative to Sabre – and UK-headquartered Travelport (which includes the Apollo, Worldspan and Galileo GDS.) All of these networks began life managing ticketing between airlines and travel agents, but today also work with travel websites, car rental firms and hotels.

Booking, please?

Sabre's earliest travel distribution network predates the internet. An airline's reservation database stores a passenger's booking information, seat selection, tickets, special requests, and other critical information about their trip. Sabre typically processes thousands of reservation updates per second on behalf of carrier customers. An airline's reservation database must be served from many availability zones.

Meanwhile, the flight shopping system generates millions of itineraries per second on behalf of travelers using mobile apps, third-party travel websites, and airline call centers. It manages 10 exabytes using Bigtable. Building a data solution to anticipate what customers might buy next is built on top of its existing data warehousing estate, which includes Teradata, Oracle and IBM.

"We have pretty much everything that you would imagine in terms of data warehousing technologies and operational data stores. For all of those existing systems we're thinking about how to initially get a feed into BigQuery," Gasparovic says.

"It was hard to do things like understanding a traveler as they went from shopping for flights to booking flights. To use that information to understand what those travelers are interested in, what their preferences are, what kind of packages of products they usually buy together. That sort of thing was hard to do because that data existed in so many different systems in so many different forms," he adds.

Sabre Labs holds on to what was offered to the traveler, creating an offer ID which is streamed into BigQuery along with what was ordered by the customer. "It's not just getting it in the same place. It's updating and modernizing the data model itself to be able to identify those things with a unique key," he says.

Being able to bring that all into the same system, start correlating it and understanding it as a whole was the first step. The next was to build machine learning models that learn from the data. "This is a really big deal for what we can provide to our customers," Gasparovic says.

But machine learning comes with a warning. It is possible to play around with machine learning forever and just not necessarily solve a real problem. Projects have to start from practical problems to solve, he says.

The Sabre Labs approach is to compress the training phase of a ML project by employing a reinforcement learning technique to very quickly try something out and watch to see what benefit it provides. "You can think of kind of like a very fancy form of A-B testing with lots of different elements being tested at the same time," Gasparovic says.

"The nice thing about that is that we don't have to do a lot of upfront training of models and trying to play around with things to get the accuracy. When we know that it is actually providing some benefit, then we can go back and do some classic machine learning: building models and doing a supervised learning process where you train a model on the data and then put that model into production and use its outputs to make predictions, but starting out with that experimentation really has been very helpful for us to know where we want to spend our time," he says.

It has already provided machine learning models that help its partner organization make better offers for travel customers.

"When we are able to put that on a website and actually see the impact, it's something that customers expect, but it's been very successful for hotels and airlines. That we've enabled this for them is still something unique in the industry," Gasparovic says.

The Google Cloud solution is built on top of Sabre's existing data warehouses, but in the long term, the plan is to consolidate data onto one system, but for "some time" Sabre is likely to be be operating both worlds with the new world layered on top of the existing world, he says.

In the meantime, Gasparovic anticipates more wins to come from the existing set-up. "We've only just scratched the surface of what we could do with machine learning and experimentation," Gasparovic says. ®