---
id: observability-red-use-metrics
title: RED / USE 메트릭 — 어떤 걸 측정할까
category: Coding
status: draft
source_trust_level: B
verification_status: conceptual
created_at: 2026-05-09
updated_at: 2026-05-09
tags: [observability, metrics, sli, slo, vibe-coding]
tech_stack: { language: "Prometheus / Grafana", applicable_to: ["Backend"] }
applied_in: []
aliases: [Rate, Errors, Duration, Utilization, Saturation, four golden signals]
---

# RED / USE 메트릭

> 측정 안 하면 운영 못 함. **RED (서비스): Rate / Errors / Duration**, **USE (리소스): Utilization / Saturation / Errors**. SRE 가 정의한 표준 출발점.

## 📖 핵심 개념
- **RED** (요청 기반 서비스): 사용자 관점.
- **USE** (리소스 기반): CPU/메모리/디스크/네트워크.
- **Four Golden Signals** (Google SRE): Latency / Traffic / Errors / Saturation.

## 💻 코드 패턴

### prom-client (Node)
```ts
import client from 'prom-client';
client.collectDefaultMetrics(); // CPU/heap/eventloop 자동

// RED — HTTP
const httpReqs = new client.Counter({
  name: 'http_requests_total',
  help: 'Total HTTP requests',
  labelNames: ['method', 'route', 'status'],
});

const httpDur = new client.Histogram({
  name: 'http_request_duration_seconds',
  help: 'HTTP request duration',
  labelNames: ['method', 'route', 'status'],
  buckets: [0.005, 0.01, 0.05, 0.1, 0.3, 0.5, 1, 2, 5],
});

app.use((req, res, next) => {
  const end = httpDur.startTimer({ method: req.method, route: req.route?.path ?? 'unknown' });
  res.on('finish', () => {
    const status = String(res.statusCode);
    end({ status });
    httpReqs.inc({ method: req.method, route: req.route?.path ?? 'unknown', status });
  });
  next();
});

app.get('/metrics', async (_, res) => {
  res.set('Content-Type', client.register.contentType);
  res.end(await client.register.metrics());
});
```

### USE — DB pool
```ts
const dbPoolUtil = new client.Gauge({
  name: 'db_pool_utilization',
  help: 'Active connections / pool size',
});
const dbPoolSat = new client.Gauge({
  name: 'db_pool_waiting',
  help: 'Connections waiting',
});

setInterval(() => {
  dbPoolUtil.set((pool.totalCount - pool.idleCount) / pool.options.max);
  dbPoolSat.set(pool.waitingCount);
}, 5000);
```

### Latency 분포 — Histogram, not Average
```promql
histogram_quantile(0.95, rate(http_request_duration_seconds_bucket[5m]))
# p95
histogram_quantile(0.99, rate(http_request_duration_seconds_bucket{route="/api/checkout"}[5m]))
```

### Error rate
```promql
sum(rate(http_requests_total{status=~"5.."}[5m]))
/
sum(rate(http_requests_total[5m]))
```

## 🤔 의사결정 기준
| 영역 | 메트릭 |
|---|---|
| HTTP API | RED — 라벨: method, route, status |
| 워커 / job queue | RED + queue length / age |
| DB / Redis | USE — connections, latency, error rate |
| 외부 API 호출 | RED — provider, status |
| Business KPI | gauge / counter (signups, paid_orders) |
| Infra (Pod CPU) | Kubernetes 자동 |

## ❌ 안티패턴
- **평균만**: 평균은 거짓말. p95 / p99 가 사용자 경험.
- **라벨 폭증**: userId / requestId 라벨 → cardinality 폭증 → 메모리 / 비용 폭사. 라벨은 enum 같은 것만.
- **histogram bucket 부적합**: 1ms~10s 인데 bucket 이 10s 단위. 의미 없음.
- **counter 와 gauge 혼동**: counter 는 monotonic. 감소 안 함.
- **리셋 시 0 dump**: counter 는 reset 알아서 처리. gauge 만 직접 set.
- **메트릭 / 로그 / trace 따로 봄**: exemplar / trace_id 로 연결.
- **알림 임계값 절대값**: 트래픽 변동에 거짓 알림. ratio + window.

## 🤖 LLM 활용 힌트
- 새 endpoint 마다 RED 자동 (middleware).
- 외부 의존성마다 별도 RED.
- p95/p99 SLO 정의 후 알림.

## 🔗 관련 문서
- [[Observability_Structured_Logging]]
- [[Observability_OpenTelemetry]]
- [[Backend_Health_Check_Patterns]]