Behind the scenes: building an 800ms image pipeline
The journal
Engineering8 min read

Hosting, queueing, and matching at speed. A look at the engineering choices that keep searches feeling instant.

Speed is a feature. When a search feels instant, people search more, explore more, and trust the result more. Hitting a sub-second round trip took deliberate engineering.

Host first, then search

Every uploaded image is pushed to a fast CDN before anything else, so the matching service always works from a public URL rather than a heavy payload. This keeps the search request small and quick.

Queue the heavy work

Matching is asynchronous: a job is submitted, then polled until it completes. That decoupling means a slow source never blocks the interface, and the front end can show progress while the work happens.

  • Upload → CDN host for a stable public URL
  • Submit an async job, poll for completion
  • Rank and merge results from multiple sources
  • Stream results back as soon as they are ready

The fastest request is the one you never have to wait on synchronously.

The result is a pipeline that feels immediate even when there is real work happening underneath.

EngineeringMore articles
keep reading