Load testing with a think time, or the numbers are fiction

A tool firing requests as fast as it can with no pause measures how fast the server can be saturated, which is not a number that corresponds to any real number of users.

# no think time: 50 connections = a permanently saturated server
$ ab -n 10000 -c 50 https://example.com/

# with think time, the result maps to something real
$ wrk -t4 -c100 -d60s --latency -s think.lua https://example.com/

# think.lua:  function delay() return math.random(2000, 8000) end

With a two to eight second pause between requests, a hundred connections approximates roughly two thousand concurrent users rather than a hundred — which is the arithmetic that makes the result quotable to somebody outside engineering. Testing a single URL is the other distortion, since a real session hits pages with different cache behaviour and a login. The number worth reporting is the concurrency at which the p99 crosses a threshold, not the peak requests per second.