simulation 04 — writes that block
A third of requests are writes that take 420ms of inline work — thumbnailing, indexing, mailing. They hold a server slot the whole time and drag the tail out for everyone, including readers. Nobody is waiting on the result.
pass when
error rate ≤ 3% · p99 ≤ 300ms · spend ≤ $460
The simulation reads best on a wider screen — the topology is drawn left to right.
Ready to run.
served
0
dropped
0
error rate
0.0%
p99
0ms
in flight
0
build
spend$100 / $460
Capacity is never free. Passing under budget is the whole exercise.