"Same Traffic, 86% Less CPU" — Valkey 9.1 Real-World Results from Imweb, a Platform Behind 1 Million Sites

Imweb upgraded its production cache running on Amazon ElastiCache from Redis 6.x/Valkey 7.2 to Valkey 9.1 in the early morning of July 10, 2026. The flagship general-purpose cache, which is repeatedly queried across the SSR path and various service logic, processed over 700 million commands over 3.5 hours.

Compared to the same time slot the previous day, overall CPU usage dropped by 86.6%, and Engine CPU dropped by 31.6%. GET latency decreased by 30.5%, SET latency decreased by 53.6%, and Engine CPU / 1M commands, adjusted for the difference in command count, also improved by 27.4%.

Imweb repeatedly queries site settings, menus, permissions, session state, and product/content display conditions along the SSR request path. On a platform with a cumulative total of 1 million sites created and 7 trillion KRW in customer transaction volume as of 2025, the cache is not just a layer for reducing DB load but a core factor determining page responsiveness and infrastructure cost.

Among the improvements from Valkey 7.2 to 9.1, the key changes that directly impacted Imweb's workload are as follows.

• I/O threading redesign: In Valkey 8.0/8.1, I/O threads now handle client read, command parsing, and response write in parallel, allowing the main thread to focus more on command execution.
• Command batching/memory prefetching: The key lookup path for multiple commands is processed in advance, reducing wait times caused by dictionary lookups and random memory access.
• Queue-based communication model: Valkey 9.1 introduces a shared queue and response queue between the main thread and I/O threads, separating work distribution from completion handling and improving load balancing.
• Memory efficiency improvements: Valkey 8.0 changed the slot-to-key linked list metadata into a per-slot dictionary structure and embedded keys directly into dictionary entries, enabling up to 20% memory savings for small key/value workloads.

This measurement is not simply a result of reduced traffic — the fact that even the command-count-adjusted engine CPU metric improved shows this is a case where Valkey 9.1's efficiency gains were confirmed in a production environment.