One reused socket served 199 of the 200 GETs behind a 12 ms mean

urllib3's pool is keyed by host, so the second call to the same host gets back the socket the first one left behind ^[5]. Run that loop against any remote HTTPS URL and the samples cluster tightly, and the field notes attribute the tightness to the pool rather than to a handler that got faster ^[17]. Averaged across 200 samples, the single request that paid DNS, TCP and TLS carries 0.5% of the mean's weight; in the reader-runnable version, with N set to 50, it carries 2% ^[2]. "Averaging cold and warm rows is not a measurement so much as a smoothing function with extra confidence," the author wrote ^[11].

The search went to the server first. Application logs, then framework middleware, then a tracing exporter that got disabled on suspicion of ten milliseconds, and none of it moved the production graph ^[8]. The process was restarted and the local resolver cache flushed, while the session stayed open between the timed calls ^[9].

"I was measuring keep-alive reuse, not the cost a cold cron job or a new pod actually pays," the author wrote ^[7]. The second harness splits the three cases that get casually mixed together: one long-lived Session for the whole loop, which is what a worker process usually does; a new Session per call, which is closer to requests.get() and to many CI checks; and the same pattern sending Connection: close, which forces the handshake back into view ^[13]. In the fresh-session path, take() builds the Session inside the loop and closes it in a finally block, so every sample there pays the full setup ^[15].

The second harness has its own limits. With N at 30, its p95 is sorted(xs)[int(0.95 * 29)], which is sorted(xs)[27]: the 28th of 30 sorted samples, third from the top, so two samples sit above the number it prints and a single outlier sets it ^[3][14]. Both harnesses call raise_for_status() inside the timed section, and the first one has no handler around it, so a non-2xx ends the run instead of showing up as a slow sample ^[4][12]. The post says the labels matter more than any absolute millisecond value from one laptop, and the text stops mid-sentence before the three rows' numbers ^[16].