Amadeus needed to accelerate its shopping engine’s response times and scale to meet surging demand from millions of online travel searches.
Amadeus migrated its shopping engine and some of its analytics workloads to Azure Cobalt 100 VMs, leveraging ARM-based architecture for improved performance.
Response times dropped by 8%, throughput increased by 20%, delivering faster, higher-quality results for travelers and partners worldwide.
As the self-described “backbone of the travel industry,” Amadeus powers the technology platforms that make modern travel possible. Its systems connect the global travel ecosystem—from travelers searching for flights, to airlines managing departures, to airports coordinating operations.
Behind this seamless experience lies a massive technical challenge: processing more than 2 billion daily transactions at peak time through Amadeus's shopping engine while keeping response times as fast as possible for customers around the world.
That challenge stems from a dramatic shift in consumer behavior. "In the '90s, all the bookings were done by travel agents," explains Didier Spezia, a platform architect at Amadeus. "Now we have numerous people on the internet trying to book trips."
That shift has created an astronomical "look-to-book ratio"—millions of searches generating relatively few actual bookings. The shopping engine must simultaneously search through all airlines for all available flights and prices, using advanced algorithms, heuristics, and brute-force computation, requiring "a lot of CPU—really a lot of CPU," Spezia says.
To handle this load, Amadeus historically ran a hybrid setup: VMware-based virtualization for traditional workloads and supporting platforms for managing containerized workloads, all from a centralized data center in Erding, Germany.
While effective, this setup was limited by global latency and made scaling challenging. Customers increasingly demanded technology deployed closer to them to integrate more easily with their own systems and to reduce latency.
The breakthrough came with the Azure Cobalt 100 VMs, the first virtual machines in Azure powered by custom Arm-based CPUs designed by Microsoft. After experimenting with Ampere processors—leading to first production deployments in early 2024—Amadeus discovered that Cobalt delivered transformative performance. By December 2024, the first Cobalt production deployment was complete, quickly followed by a full rollout—a remarkably swift transition for a workload of this scale.
Cobalt's architecture maps one vCPU (virtual CPU) to a single full core, different from the simultaneously multithreaded processors typically seen on similar Intel and AMD-based VMs. For Amadeus's compute-heavy shopping engine, this design proved ideal for transactional throughput. This shopping engine is using Red Hat Enterprise Linux containers managed by Red Hat OpenShift.
The migration required significant coordination but also delivered pleasant surprises. Because Arm uses a different architecture, Amadeus's C++ shopping application required recompilation and extensive testing across multiple teams. However, the actual code changes proved lighter than expected. Antoine Collier, Principal Engineer who coordinated the migration, says the team found good compatibility with their open-source libraries, requiring minimal code adaptation.
Once the foundational Arm work was complete, the subsequent move from Ampere to Cobalt required no additional compilation. "We took the binaries, we shipped to the new VMs and it worked," says Luc Choubert, Vice President of platform engineering.
Amadeus also migrated some Azure Databricks big data analytics workloads to Azure Cobalt 100 VMs. That transition was seamless; Spark jobs ran without code changes, thanks to Arm compatibility. This move improved throughput, enabling faster analytics and fueling AI innovation for internal and customer-facing workloads.
Today, 90% of Amadeus's shopping engine runs on Cobalt 100 VMs, with 95% of production workloads in Azure. The company is on track to complete its full migration from on-premises data centers. The platform spans multiple European Azure regions where Ireland and the Netherlands are main hubs, with France, Sweden and Germany as satellites.
The results speak for themselves: response times for the shopping engine dropped by 8%, and throughput increased roughly 20%, improving the price-performance of the solution. This performance gain enabled Amadeus to maximize answer quality within their response time envelope, a critical metric when it comes to attracting and retaining customers.
"We are constantly compared to our competitors," explains Agnès Penaud, Director of Shopping and Pricing Engine R&D. "Those who are the most rapid and best in quality win the bookings."
Looking ahead, Amadeus plans to explore Cobalt 100 VMs for additional workloads, including Couchbase databases and Java-based cloud-native applications, while expanding regionally and deepening AI exploration for both internal efficiency and customer-facing products.
Discover more about Amadeus on Facebook, Instagram, LinkedIn, X/Twitter, and YouTube.
“We are constantly compared to our competitors. Those who are the most rapid and best in quality win the bookings.”
Agnès Penaud, Director of Shopping and Pricing Engine R&D, Amadeus
Follow Microsoft