Write a Realistic Load Test Scenario

1. Define an arrival-rate scenario with a ramp, a plateau, and explicit thresholds.

import http from 'k6/http';
import { check, sleep } from 'k6';

export const options = {
  scenarios: {
    browse: {
      executor: 'ramping-arrival-rate',
      startRate: 20,
      timeUnit: '1s',
      preAllocatedVUs: 50,
      maxVUs: 300,
      stages: [
        { target: 200, duration: '2m' },
        { target: 200, duration: '5m' },
        { target: 0, duration: '1m' },
      ],
    },
  },
  thresholds: {
    http_req_duration: ['p(95)<300', 'p(99)<800'],
    http_req_failed: ['rate<0.01'],
  },
};

Expected output: 200 requests per second for five minutes; the run fails if p95 exceeds 300 ms or errors exceed 1 percent.

2. Script a short user journey with think time instead of hammering one endpoint.

export default function () {
  const list = http.get(`${__ENV.BASE_URL}/api/products?limit=20`);
  check(list, { 'list 200': (r) => r.status === 200 });
  const id = list.json('data.0.id');
  const detail = http.get(`${__ENV.BASE_URL}/api/products/${id}`);
  check(detail, { 'detail 200': (r) => r.status === 200 });
  sleep(Math.random() * 2 + 1);
}

Expected output: Each virtual user browses a list, opens one detail page, and waits 1 to 3 seconds, matching real sessions.

3. Spread requests over a realistic identifier set so caches and indexes behave like production.

import { SharedArray } from 'k6/data';
const ids = new SharedArray('product ids', () => JSON.parse(open('./ids.json')));
const id = ids[Math.floor(Math.random() * ids.length)];

Expected output: Traffic hits thousands of distinct products instead of one row that stays in the buffer cache.

4. Run against staging and record the summary with the release candidate.

k6 run -e BASE_URL=https://staging.example.com --summary-export=reports/load-2026-10-21.json load/products.js

Expected output: Thresholds pass with p95 around 210 ms, p99 around 540 ms, and an error rate under 0.5 percent.

5. Compare the summary with the previous release instead of reading a single number in isolation.

node -e "const a=require('./reports/load-2026-10-21.json'); const b=require('./reports/load-2026-09-15.json'); console.log('p95', b.metrics.http_req_duration.values['p(95)'], '->', a.metrics.http_req_duration.values['p(95)'])"

Expected output: p95 195 -> 210, a 7 percent regression that is reviewed and either accepted or fixed before release.

Problem

Thresholds fail only in the first minute

Solution

Caches and connection pools are cold. Add an unmeasured warmup stage, then start the ramp so the plateau reflects steady state.

Problem

The load generator runs out of CPU before the service does

Solution

Watch k6 CPU and the dropped_iterations metric. Distribute the test across multiple instances or use a managed runner when the target rate is high.

Problem

Results vary wildly between runs

Solution

Pin the dataset, the instance type, and the time of day. Run three times and compare medians; a single run is an anecdote, not a baseline.