Recombee scales real-time personalization with OVHcloud® Bare Metal

Recombee's infrastructure delivers up to 100,000 real-time recommendation results per second during major global media and sporting events. Operating across five regions, the company manages more than 1,000 dedicated servers and 6,000 application instances, an estate that includes database servers storing over 700TB of data, event queues, request routers, and machine learning (ML) workers. A lean team of around four infrastructure engineers keeps this global footprint running. Recombee's customers include major publishers, video streaming services, and e-commerce companies, including The Telegraph, DAZN, and 9GAG.

Challenges:

Finding a global infrastructure partner strong enough for a resource-heavy ML platform

Recombee built its earliest infrastructure on dedicated bare metal in Europe, choosing a cost-efficient setup that provided the computational performance and control needed to support its demanding recommendation workloads. This infrastructure served Recombee well through its early stages of product development and growth. As the company expanded internationally and its workloads increased, however, new requirements emerged around storage capacity, memory flexibility, and access to more powerful infrastructure for machine-learning workloads. 

Recombee also needed to deploy infrastructure closer to customers in new regions, where the physical distance to its existing European cluster would otherwise result in unacceptable latency.

Australia proved particularly challenging. Recombee set out to build a local Australian cluster to provide low-latency service to customers in Australia and the wider region, including e-commerce customer Temple & Webster. However, the local infrastructure provider Recombee initially selected struggled to deliver the reliability and network performance required.

The most challenging region for us has been Australia. We had difficulty finding a provider that could offer even a reasonably reliable service, with networking at a reasonable speed. There's always a trade-off between price and quality, but the provider we had in Australia was really terrible, so we decided to migrate to OVHcloud®.
- Tomas Barton, Head of Infrastructure, Co-Founder
 

The experience highlighted the importance of finding an infrastructure partner that could provide consistent performance and reliability across every region in which Recombee operated. Recombee ultimately migrated its Australian infrastructure to OVHcloud as part of its broader approach to building reliable, low-latency regional deployments.

Recombee's list of requirements for a global infrastructure partner included:

• Reliable server performance and network stability in every region, not just Europe;
• Scalable storage capacity and memory allocation to keep pace with growth;
• More powerful ML worker instances to support Recombee's machine learning pipeline;
• Predictable, cost-effective pricing that would hold up at global scale.

Solution:

The bare metal foundation for real-time recommendations at global scale

Recombee had already expanded into Canada with OVHcloud®, which established a strong baseline for the region. Building on that experience, Recombee standardized on OVHcloud Advance Dedicated Servers as the foundation of its global infrastructure. High bandwidth capacity and flexible configurations gave Recombee's evolving ML architecture the raw performance and reliability its real-time recommender systems need.

Server nodes typically run around 16 CPU cores and 256GB of RAM, scaling to 40 or more cores depending on the region. Recombee uses vRack® Private Networking to securely interconnect its infrastructure across regions, along with failover IP addresses for seamless routing.

Recombee's architecture runs on two entry points. The Recommendation API delivers low-latency results through high-performance GPU and ML worker nodes, using Recombee's proprietary ReQL (Recombee Query Language) for complex filtering. The Interaction & Catalog API manages high-volume data ingestion, processing user interactions and catalog updates asynchronously through message queues and time-series data stores, so live query performance never suffers.

Since we expanded and started using OVHcloud®, we noticed that OVHcloud is much more reliable. The hardware is more performant, and the quality of the networking is substantially better than what we were used to with our previous provider. That's one big difference.
- Tomas Barton, Head of Infrastructure, Co-Founder

Migrating to OVHcloud in Australia, and standardizing across Canada, Australia, and the US West, went smoothly. Careful planning by Recombee's technical team meant the transition was completed without downtime or technical issues. The OVHcloud interface provided a clear, functional environment for managing infrastructure across all three regions, and that usability, combined with predictable platform behavior, let Recombee's team execute its migration plan exactly as designed.

Results:

Faster, cheaper, and more reliable recommendations

Cutting costs without cutting performance

Standardizing on OVHcloud bare metal cut Recombee's infrastructure costs by 60% to 80%, by eliminating markups and unpredictable egress fees. Removing virtualization overhead helps Recombee maintain global latency under 200ms during peak loads of 100,000 requests per second, and its multi-region configuration allows for global system updates in minutes. The infrastructure gains translate directly into customer results. For The Telegraph, Recombee's recommendation boxes on article detail pages have delivered a 35% increase in click-through rate.

Key outcomes:

• 60% to 80% lower infrastructure costs by eliminating markups and unpredictable egress fees. 
• Global latency under 200ms at peak loads of 100,000 requests per second
• Multi-region system updates completed in minutes

 OVHcloud® is a registered trademark of OVH SAS. vRack® is a registered trademark of OVH SAS.

Ready to Get Started?

Contact us