Benchmarking a cold endpoint: why splitting latency into phases matters

A developer found that a single averaged latency figure hid two distinct behaviors: a tight median and a much slower tail that changed with the time of day. Recording each request's wall time as one number mixed name resolution, TCP connect, TLS handshake, and server response, making it impossible to tell network delay from application slowness. The author instead tracked connect, TLS, send, time-to-first-byte, and total durations separately, which distinguishes a slow server from an idle connection that needed waking up. A curl command with timing format flags was used to confirm the breakdown quickly, and a standard-library Python probe was written to log each phase per attempt. The key lesson is to identify which phase is being measured before optimizing, since a profiler cannot diagnose network delays.