4.1 配套代码：k6 进阶用法与压测机自查

// loadtest/multi-scenario.js
import http from 'k6/http';
import { check } from 'k6';
import { Trend, Rate, Counter } from 'k6/metrics';

// 自定义指标（对应第 1 章的四层指标）
const readLatency = new Trend('read_latency_ms', true);
const writeLatency = new Trend('write_latency_ms', true);
const businessErrors = new Counter('business_errors');
const okRate = new Rate('ok_rate');

export const options = {
  scenarios: {
    // 读路径：占 90% 流量，开环恒定
    read_path: {
      executor: 'constant-arrival-rate',
      rate: 450, timeUnit: '1s', duration: '10m',
      preAllocatedVUs: 500, maxVUs: 3000,
      exec: 'readScenario',
      tags: { path: 'read' },
    },
    // 写路径：占 10% 流量，开环较低速率
    write_path: {
      executor: 'constant-arrival-rate',
      rate: 50, timeUnit: '1s', duration: '10m',
      preAllocatedVUs: 200, maxVUs: 1000,
      exec: 'writeScenario',
      startTime: '30s',              // 错开启动，避免冷启动叠加
      tags: { path: 'write' },
    },
  },
  thresholds: {
    'http_req_failed{path:read}': ['rate<0.01'],
    'http_req_duration{path:read}': ['p(99)<200'],
    'read_latency_ms': ['p(99)<200'],
    'write_latency_ms': ['p(99)<500'],
    ok_rate: ['rate>0.99'],
    business_errors: ['count<100'],
  },
  discardResponseBodies: true,
  summaryTrendStats: ['avg', 'min', 'med', 'p(90)', 'p(95)', 'p(99)', 'p(99.9)', 'max'],
};

function zipf(max) {
  return Math.max(1, Math.floor(max * Math.pow(Math.random(), 3)) + 1);
}

export function readScenario() {
  const res = http.get(`${__ENV.BASE_URL}/orders/${zipf(1_000_000)}`, {
    tags: { name: 'GET /orders/{id}' },      // ⚠️ 路由模板，不是具体 URL
  });
  readLatency.add(res.timings.duration);
  const ok = check(res, { 'read ok': (r) => r.status === 200 });
  okRate.add(ok);
  if (!ok) businessErrors.add(1);
}

export function writeScenario() {
  const payload = JSON.stringify({ userId: Math.floor(Math.random() * 1000) + 1, amount: 100 });
  const res = http.post(`${__ENV.BASE_URL}/orders`, payload, {
    headers: { 'Content-Type': 'application/json', 'X-Load-Test': 'true' },
    tags: { name: 'POST /orders' },
  });
  writeLatency.add(res.timings.duration);
  const ok = check(res, { 'write ok': (r) => r.status === 201 });
  okRate.add(ok);
  if (!ok) businessErrors.add(1);
}

#!/usr/bin/env bash
# tools/check-load-generator.sh <SUMMARY_JSON> <EXPECTED_RATE>
#
# 判断"压测机是不是瓶颈"——这是数据分析前的第一道关。
set -uo pipefail

SUMMARY="${1:?usage: check-load-generator.sh <summary.json> <expected_rate>}"
EXPECTED="${2:?}"

echo "═══ 压测机自查 ═══"
echo

# ① 到达率偏差（最重要）
python3 - "$SUMMARY" "$EXPECTED" <<'PY'
import json, sys
summary, expected = sys.argv[1], float(sys.argv[2])
m = json.load(open(summary))["metrics"]
actual = m["http_reqs"]["rate"]
diff = actual / expected - 1
print(f"① 到达率检查")
print(f"   设定 : {expected:>10.1f} req/s")
print(f"   实际 : {actual:>10.1f} req/s")
print(f"   偏差 : {diff:>+10.1%}")
if abs(diff) > 0.05:
    print("   ❌ 偏差 > 5% —— 压测机可能是瓶颈，服务端数据不可用")
    print("      排查方向：preAllocatedVUs 不足 / 压测机 CPU 打满 / 端口耗尽")
else:
    print("   ✅ 通过")
print()

# ② 分布形态检查（P99/P50 比值）
d = m["http_req_duration"]
p50, p99 = d.get("med", 0), d.get("p(99)", 0)
if p50 > 0:
    ratio = p99 / p50
    print(f"② 分布形态检查")
    print(f"   P50 {p50:.1f} ms, P99 {p99:.1f} ms, 比值 {ratio:.1f}")
    if ratio < 2:
        print("   ⚠️  P99/P50 < 2 —— 分布过于'干净'，怀疑闭环模型导致尾延迟被低估")
    else:
        print("   ✅ 分布形态符合重尾特征")
print()

# ③ 错误率
failed = m["http_req_failed"]["rate"]
print(f"③ 错误率检查")
print(f"   {failed:.2%}")
if failed > 0.01:
    print("   ❌ > 1% —— 延迟数据可能因剔除失败请求而虚低")
else:
    print("   ✅ 通过")
PY

echo
echo "═══ 压测机 OS 层检查 ═══"

# ④ 临时端口与连接状态（TIME_WAIT 堆积是常见瓶颈）
if command -v ss > /dev/null 2>&1; then
  echo "④ 连接状态："
  ss -s | head -5
  TW=$(ss -tan 2>/dev/null | grep -c TIME-WAIT || echo 0)
  echo "   TIME_WAIT 数量: $TW"
  [ "$TW" -gt 20000 ] 2>/dev/null && echo "   ⚠️  TIME_WAIT 过多，考虑开启 tcp_tw_reuse 或增加连接复用"
fi

# ⑤ 压测机负载
echo
echo "⑤ 压测机负载（如果你能看到）："
echo "   运行 'top' 或 'pidstat -u 1' 观察 k6 进程的 CPU"
echo "   如果 k6 进程 CPU 接近 100%（单核），说明压测客户端是瓶颈"

# 工具 1：k6（开环）
k6 run --summary-export=k6.json loadtest/profile-constant.js

# 工具 2：wrk2（恒定速率 + HdrHistogram 修正）
wrk2 -t4 -c100 -d60s -R500 --latency http://127.0.0.1:8080/orders/1

# 工具 3：wrk（闭环）—— 故意用它来看差异
wrk -t4 -c100 -d60s --latency http://127.0.0.1:8080/orders/1

# 对比三者的 P99
echo "k6   : $(python3 -c "import json;print(json.load(open('k6.json'))['metrics']['http_req_duration']['p(99)'])") ms"