Congestion collapse
what retries do to a service that is already late
What people expectThat backing the retries off undoes the collapse. The audience expects the system to be a function of its current input, and it is a function of its history — the queue is by then deep enough to time out its own contents.
The numbers
- service time ms
- 40 ms
- timeout ms
- 120 ms
- capacity rps
- 12,000 req/s
- offered load rps
- 10,500 req/s
- queue limit
- 4,000 requests
- retry budget
- 10 × base load
- seed
- 20260729
The mechanism is documented; these particular values are authored, chosen to put the interesting threshold somewhere you can see it. The simulation is deterministic and seeded — the same spec renders the same frames on any machine.