Developer Discovers 48-Hour API Debug Was a Client-Side Keep-Alive Measurement Error

A developer spent 48 hours investigating a mismatch between a locally measured 12ms API latency and a much higher production p95 response time. The root cause turned out to be a Python requests.Session reusing a single TCP and TLS connection across 200 sequential calls, meaning subsequent requests never paid the cost of a handshake. Because the benchmark loop never closed the session between timed calls, the recorded mean reflected warm keep-alive reuse rather than the true cold-start cost a new pod or cron job would experience. The developer traced the issue by printing individual samples and noticing the first call was several times slower than all subsequent ones. The findings highlight how averaging cold and warm latency samples produces misleadingly optimistic numbers that can end up embedded in design documents.