Designing an Efficient Real-Time Feature Store at Zomato

Zomato is one of India's leading food delivery and restaurant discovery platforms, serving millions of customers across hundreds of cities. Its real-time Profile Store powers recommendation and personalization ML models by storing and serving behavioral, historical, and contextual features for customers, restaurants, dishes, cuisines, and delivery partners at massive scale.

To support growing traffic and increasingly demanding ML feature-data workloads, Zomato redesigned its AWS-based real-time feature store to handle approximately 40M RPM while significantly reducing infrastructure costs, improving latency, and scaling efficiently through optimized serialization, intelligent caching, and concurrency-aware request handling.

In this session, we will share the architectural decisions and engineering techniques behind the redesign, including Protocol Buffer optimization, DynamoDB storage and network efficiency improvements, Redis cache optimization, and request coalescing to eliminate duplicate concurrent requests. We will also discuss the key architectural trade-offs, operational challenges, and lessons learned from running a production-grade, high-throughput, low-latency feature store at scale.

Key Outcomes:

DynamoDB

Reduced read and write costs by approximately 30%.

Achieved nearly 4x payload size reduction, improving storage and data transfer efficiency.

ElastiCache (Redis)

Reduced infrastructure footprint from 354 nodes to 196 nodes, delivering approximately 45% infrastructure cost savings.

Lowered maximum Redis CPU utilization from 60% to 35% while improving latency by 2-3 ms.

We believe this session would be valuable for engineers and architects building large-scale, low-latency ML feature stores, offering practical insights from operating a production system at very high throughput.

Thank you for considering our submission. We would be happy to provide any additional information if needed.