Performance & Scaling
Nothing looks wrongLook again
In performance & scaling, the system that holds today is already deciding where it will break.The load that breaks it behaves nothing like staging.
Under load
The ceiling is already in the architecture
A system rarely slows down gradually – it holds fine until one shared constraint saturates, and then it falls off a cliff
The signs it won't survive contact with production
- 01p99 already driftingAverages stay flat while the p99 has been climbing for weeks, the first place saturation shows up
- 02One shared hot pathEvery request funnels through a single hot row or table nobody has yet had a reason to shard
- 03Latency tracks concurrency, not dataResponse times follow the number of in-flight requests rather than query size – a queue forming out of sight
- 04Connection pool near its ceilingThe database pool already sits at 80% under normal load, with no headroom left for a spike or a slow query
- 05Staging never sees real concurrencyLoad is validated one request at a time, so contention that only appears in parallel never surfaces before production