KakaoTalk Reduces Servers from 256 to 180 by Building Kubernetes-Based Redis Cache Farm

The KakaoTalk Talk Server team transitioned its existing manually operated cache system based on Memcached to a 'cache farm' leveraging Kubernetes and Redis. This move was aimed at addressing the high memory waste and manual failure recovery burden associated with handling data access traffic exceeding 4M/s.

The new architecture uses StatefulSets to flexibly deploy Redis pods based on memory capacity (1GB, 4GB, 16GB). To minimize network overhead, hostNetwork was used, but to resolve the resulting client abstraction issues, Redis Sentinel and an in-house operational tool 'Barn' were introduced. Barn manages the Desired State and collaborates with Sentinel to automatically assign new nodes and restore client connections upon failure.

Through this transition, the number of physical servers was reduced by over 30%, from 256 to 180, and overall memory availability improved from 28% to 48%. Additionally, manual tasks during late-night hours and communication costs with the infrastructure team were significantly reduced, securing operational efficiency and system resilience.