How we cut p95 latency by 60% without adding servers

We doubled traffic and the p95 still dropped 60% — on the same boxes. Here's where the time was actually going.

Profile before you guess

Every latency investigation starts the same way: stop guessing, start measuring. A flamegraph of the hot path showed the bottleneck was nowhere near where the team assumed — it was serialization, not the database.

Cache the shape, not just the data

The biggest win came from caching pre-computed response shapes in Redis rather than re-deriving them per request. The data was already fast to fetch; assembling it wasn't.