Back to Customers
Zalando

Zalando

Europe's largest online fashion platform powers real-time personalization across 25 countries and 50 million customers with Hopsworks.

Visit website

How Zalando serves real-time features at scale with Hopsworks

<10ms
Online Feature Serving Latency
Multi-AZ
High Availability Replication
Dedicated
Clusters for Critical Workloads

"We ran extensive benchmarks comparing both setups and saw no noticeable performance degradation on Kubernetes. The reduced manual effort and built-in autoscaling were major wins for us. Performance remained on par while operational efficiency improved significantly."

Real-Time Feature Serving for Personalization at Scale

Zalando is Europe's largest online fashion platform, operating in over 25 countries and serving more than 50 million customers. To power real-time personalization use cases such as recommendations, Zalando needed a scalable, low-latency, and highly available feature serving platform that could support many teams and mission-critical applications.

The Aim

As Zalando scaled its machine learning platform and customer-facing use cases, several challenges emerged:

  • Feature data was fragmented across teams, leading to data silos
  • Inconsistent features between training and production environments
  • Limited discoverability and reusability of features across teams and departments
  • Difficulty ensuring low-latency feature access for real-time customer-facing applications
  • High operational overhead in maintaining feature serving infrastructure
  • Risk of large blast radius during incidents due to shared infrastructure
  • Limited scalability and resilience in earlier EC2-based deployments
  • Manual operations and configuration drift across environments
  • Challenges meeting strict availability and latency SLOs for critical use cases

Zalando needed a centralized feature platform that could provide strong isolation, high availability, and predictable performance while supporting collaboration across the organization.

Why Hopsworks?

Zalando adopted the Hopsworks Feature Store as a centralized, real-time feature platform to support both online and offline ML workloads.

With Hopsworks, Zalando was able to:

  • Centralize feature storage to reduce data silos and improve consistency across training and serving
  • Enable feature discovery, versioning, and reuse across multiple teams and projects
  • Serve features with strict low-latency requirements for customer-facing applications
  • Achieve high availability through multi-availability zone replication
  • Reduce blast radius by isolating critical projects in dedicated clusters
  • Manage infrastructure through APIs and Infrastructure as Code
  • Transition from EC2-based deployments to Kubernetes (EKS) for better isolation and self-healing
  • Leverage RonDB for scalable, low-latency online feature serving
  • Integrate event-driven pipelines using Kafka for feature freshness
  • Consolidate metadata, access control, and governance across federated feature stores

Results

Low-Latency Feature Serving at Scale

  • Sub–10 millisecond latency for latency-sensitive personalization use cases
  • Predictable performance under high request volumes
  • Real-time feature freshness for more accurate recommendations

Improved Reliability and Availability

  • High-availability feature serving with multi–availability zone replication
  • Reduced blast radius through project-level cluster isolation
  • Self-healing services on Kubernetes, minimizing operational incidents

Centralized Feature Governance

  • Single source of truth for features across teams and departments
  • Centralized metadata, versioning, and access control
  • Easier collaboration and reuse of high-quality, production-ready features

Faster Scaling with Lower Operational Overhead

  • Automated horizontal scaling based on response time, request rate, and CPU usage
  • Seamless capacity expansion without manual intervention
  • Infrastructure managed via APIs and Git-based configuration
  • Elimination of manual EC2 operations such as SSH access and disk resizing
  • Faster rollouts and upgrades using rolling Kubernetes deployments

Ready to join them?

See how Hopsworks can transform your ML operations.

Start NowContact Sales