This article describes how Booking.com built an ultra-low latency feature platform using Amazon ElastiCache to serve millions of real-time ML predictions per minute.

• Achieves p99.9 latency below 25ms at 200,000 requests per second
• Delegates feature computation to external systems like Snowflake, Flink, Spark
• Uses Kafka-based ingestion pipeline for scalability and reliability
• Implements per-use-case ElastiCache clusters for workload isolation
• Supports two storage layouts: Key-JSON and Key-Kryo serialization
• Schema-driven design enables data integrity and schema evolution
• Self-service configuration management through Git and CI/CD pipelines
• REST API service deployed on Amazon EKS for high availability
• Offline store integration with Kafka and Snowflake for analytics
• Ranking system migration delivered $3M annual savings with 82% cost reduction

Booking.com's ElastiCache-based feature store enables 22+ teams to deploy real-time ML pipelines with ultra-low latency, reducing infrastructure costs while maintaining strict performance requirements.