How Netflix’s EVCache Powers Global Low‑Latency Caching Across Regions

EVCache Overview

Netflix runs hundreds of stateless micro‑services that keep all state in caches or persistent stores. EVCache, built on Memcached and optimized for cloud environments, provides a low‑latency, highly reliable caching layer that can handle tens of billions of objects across thousands of servers.

EVCache is open‑source and has been in production for over five years, processing about 2 × 10¹³ requests per day. It operates globally on AWS regions (Virginia, Oregon, Ireland) and relies on asynchronous replication via Kafka to keep caches consistent across regions.

When a client calls set(), the local EVCache instance stores the value, writes only the key and metadata to a Kafka replication queue, and a replication service reads the message, fetches the value from the local cache, and issues a SET to the remote region’s cache proxy. GET operations read from the local cache only.

The replication design avoids sending full data over Kafka, reducing bandwidth and latency, and uses persistent TCP connections and batch‑sending to improve throughput. Typical end‑to‑end replication latency is under 400 ms for 99 % of requests, handling up to one million RPS globally.

Current challenges include scaling Kafka partitions, handling server failures that increase latency, and monitoring message loss or slow partitions. Ongoing work focuses on improving Kafka scalability, fault‑tolerant proxy behavior, and better alerting for replication delays.