Vercel CDN: indexed metadata shards cut P99 lookup latency 91%

On September 10, 2026, Tim Caswell, Steven Salat, and Luba Kravchenko described how Vercel cut CDN metadata lookup latency. The CDN runs on average over 80 million routing instructions per second, and metadata lookups decide which paths exist and how to serve them. The team moved from per-path metadata objects to bounded indexed shards of about 200 KB, using JSONL plus inline indexes so one fetch warms many paths while lookups parse only the needed record.

Production measurements on August 5-12, 2026 traffic show P99 metadata lookup latency falling from 215.8 ms to 19.1 ms (91% lower) and the average from 8.59 ms to 1.81 ms (79% lower). Deployments built after July 17, 2026 already use the new shards; older deployments need a redeploy. The Build Output API contract is unchanged, and the change also saved about 16.6 seconds of redundant metadata upload work, making the deploy step about 10% faster overall. That routing is the same Vercel deploy path you already use in Chapter 1, and a redeploy can pick up the platform change without app code changes.

Read original: How we cut CDN metadata lookup latency by 91% — Vercel

Related in the book: Chapter 1