# How we cut CDN metadata lookup latency by 91% By Steven Van · 2026-09-10 The CDN now groups routing metadata into 200 KB shards instead of per-path lookups, based on production tests run in August 2026. [Vercel]() rebuilt how its CDN looks up routing metadata for incoming requests, cutting P99 lookup latency by 91%, from 215.8 ms to 19.1 ms, based on production traffic measured August 5-12, 2026. The CDN used to cache metadata one path at a time, so every new deployment created fresh cache keys and a fresh round of cache misses, a recurring cost for large sites with hundreds of thousands of paths that deploy frequently. Vercel now groups path metadata into bounded "shards" of about 200 KB, each with an inline index that lets the CDN binary-search for a single path's metadata without parsing the rest of the shard. One fetch now warms every path in a shard instead of just one. Average lookup latency dropped 79%, from 8.59 ms to 1.81 ms, and standard deviation fell 58%, from 44.9 ms to 19.0 ms. Vercel [tested the change]() offline against a harness that compared old and new lookups for every path, then ran both lookups in production behind a feature flag for several weeks before switching over. [![]()]() ![Vercel]() Vercel The platform for frontend developers — deploy, preview, and scale web apps and AI agents with zero config. [View Vercel →]() Sources [Original source]() [Read How we cut CDN metadata lookup latency by 91% on Creators Toolbox]() --- Canonical source: https://creatorstoolbox.com/blog/vercel-how-we-cut-cdn-metadata-lookup-latency-by-91