Vercel Reduces P99 Latency by 91% with Shard-Based Metadata Lookup

Vercel introduced a Shard-based Group Fetching architecture to address scalability limits in its CDN environment, which processes over 80 million routing directives per second, replacing the previous method of individual metadata lookups per path. This change reduced cache misses during large-scale deployments and decreased P99 metadata lookup latency by 91% in production.

Shard-Based Lookup Architecture

The previous approach generated independent cache keys for each path, causing frequent cache misses during deployments. The new approach groups metadata for multiple paths into Shard files, allowing all paths within a shard to be cache-warmed with a single fetch. Each shard includes an inline index to identify individual path locations and is optimized to a size of 200 KB to maintain LRU cache hit rates while controlling transfer costs.

Performance Improvements and Build Optimization

Production measurements showed significant improvements in metadata lookup latency.

• P99 Latency: Reduced by 91%, from 215.8 ms to 19.1 ms
• Average Latency: Reduced by 79%, from 8.59 ms to 1.81 ms
• Standard Deviation: Reduced by 58%, from 44.9 ms to 19.0 ms

Additionally, the build pipeline was optimized by omitting unnecessary per-path metadata uploads and integrating routing group information into the manifest. This resulted in overall deployment stages being approximately 10% faster, with deployments containing high volumes of metadata estimated to be accelerated by about 25%.

Compatibility and Rollout

The new format does not change the existing Build Output API contract, making it transparent to framework developers. It is automatically applied to deployments built after July 17, 2026, while previous deployments can benefit from improved lookup speeds through redeployment. Furthermore, reliability was ensured by fixing emoji handling bugs in legacy encoding (RFC 2047) by switching to pure UTF-8 storage.