Scaling redirect handling to infinity on Vercel

When the scale of redirects is small, there's no problem, but once they grow to the scale of millions, latency and cost become core challenges for the system. With existing Routing rules, the amount of work per request increases linearly as rules grow, and Middleware causes latency by running additional code on every request.

To solve this, Vercel introduced a new approach combining Bloom filter and Sharding. A Bloom filter is a structure that probabilistically checks whether data is included in a set, filtering out most requests that don't need a redirect very quickly, minimizing latency.

To solve the memory problems that arise as data scale grows, Sharding technology was applied. Instead of managing all the data as one giant file, paths are hashed and distributed across multiple smaller Shards.

When a request comes in, the Bloom filter is checked first, and only when there's a possible match is the specific corresponding shard fetched and parsed. This reduces the memory burden on the process and efficiently utilizes external storage and the file system cache.