16 · Measure, protect, and pay for the service

Deep systems understanding · Study map 77 of 85

Load testing: recreate the traffic, not just the average

Suggested preparation: Observability: connect the user symptom to the machine

Picture it

Testing a bridge by walking across it alone tells you nothing about rush hour. Real users arrive in sudden, unpredictable waves, not a steady, polite single-file line. Open-loop testing unleashes this realistic flood, revealing the breaking points where queues pile up and the structure begins to groan under the weight.

What to understand

  • Open versus closed loop, trace replay, warm/cold caches
  • Arrival process, length distributions, steady state, errors, and tail metrics

Follow the references

  1. 1 · Documentation

    vLLM · Serving benchmark reference ↗

    Inspect request-rate, burstiness, dataset, and latency options. Read the current version’s metric definitions.

Make it concrete

Sweep offered load from idle through overload. Count all arrivals, rejections, timeouts, and successful outputs; plot goodput and tails alongside throughput.

You understand it when…

Why is reporting only completed-request latency misleading when many requests timed out?

This is a study map. Follow the reference for the full explanation; use the question to test your mental model. Browse the reference shelf.