asepharyana-hub-guide/skills/monitoring/SKILL.md

---
name: monitoring
description: Monitoring and observability best practices — Prometheus, Grafana, alerts, dashboards, uptime monitoring, and incident response. Use when setting up monitoring infrastructure, designing dashboards, defining alerts. Triggers from project files and configuration, not just keyword matching."
---

# Monitoring Best Practices

## Core Concepts

- **SLI** (Service Level Indicator) — what you measure (latency, error rate, uptime).
- **SLO** (Service Level Objective) — target value (p99 < 500ms, error rate < 0.1%).
- **SLA** (Service Level Agreement) — contractual commitment. Usually looser than SLO.

**Rule:** Set SLOs tighter than SLAs so you detect problems before customers do.

## Prometheus Setup (this repo)

```yaml
# Prometheus auto-discovers containers with this label
docker service create \
  --label prometheus.io/scrape=true \
  --label prometheus.io/path=/metrics \
  --label prometheus.io/port=8080 \
  ...
```

### Key Metrics to Export

Every service should expose a `/metrics` endpoint:

```prometheus
# Service metrics
http_requests_total{method="GET", path="/users", status="200"}
http_request_duration_seconds{quantile="0.5", quantile="0.9", quantile="0.99"}
http_requests_in_flight
errors_total{type="validation", type="database", type="auth"}

# Traefik metrics (auto-exposed via `--metrics.prometheus=true`)
traefik_service_request_duration_seconds{service="...", quantile="..."}
traefik_service_requests_total{service="...", code="..."}

# System
process_cpu_seconds_total
process_resident_memory_bytes
process_open_fds
```

### Recording Rules
```yaml
groups:
  - name: service
    rules:
      - record: service:error_rate_5m
        expr: rate(errors_total[5m]) / rate(http_requests_total[5m])
      - record: service:latency_p99_5m
        expr: histogram_quantile(0.99, rate(http_request_duration_seconds_bucket[5m]))
```

## Grafana Dashboards

### Dashboard Components

| Panel | Metric | Good |
|-------|--------|------|
| **RPS** | `rate(http_requests_total[5m])` | Matches traffic patterns |
| **Error rate** | `service:error_rate_5m` | < 1% |
| **Latency** | `service:latency_p99_5m` | < 500ms |
| **CPU** | `process_cpu_seconds_total` | < 80% sustained |
| **Memory** | `process_resident_memory_bytes` | Steady, no leaks |
| **Active connections** | `http_requests_in_flight` | < configured max |
| **Open file descriptors** | `process_open_fds` | < 50% of limit |

### Dashboard Design Rules
- **Single pane of glass** — most important metrics visible without scrolling.
- **Red/yellow/green thresholds** — at a glance status.
- **Time range controls** — last 15m, 1h, 6h, 1d, 7d.
- **Template variables** — select by service, host, environment.
- **Annotations** — mark deployments, config changes on timeline.

## Alerting Rules

```yaml
groups:
  - name: alerts
    rules:
      - alert: HighErrorRate
        expr: service:error_rate_5m > 0.01
        for: 5m
        labels:
          severity: critical
        annotations:
          summary: "{{ $labels.service }} error rate is {{ $value | humanizePercentage }}"

      - alert: HighLatency
        expr: service:latency_p99_5m > 1.0
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "{{ $labels.service }} p99 latency is {{ $value }}s"

      - alert: ServiceDown
        expr: up{job="~.*"} == 0
        for: 1m
        labels:
          severity: critical
        annotations:
          summary: "{{ $labels.job }} has been down for >1 minute"
```

### Severity Levels

| Level | Response | Time to Acknowledge |
|-------|----------|---------------------|
| **P1 (Critical)** | Pages on-call | 5 min |
| **P2 (High)** | Alerts team during business hours | 30 min |
| **P3 (Medium)** | Ticket, next business day | 8 hours |
| **P4 (Low)** | Backlog, no deadline | N/A |

## Uptime Monitoring

- **Synthetic checks** — test critical user journeys every minute.
- **SSL certificate expiry** — alert when <30 days remaining.
- **Blackbox monitoring** — external service checking your endpoints.
- **Heartbeat** — cron job pings a Dead Man's Switch — if it stops, on-call is paged.

## Incident Response

### Runbook Template
```markdown
# Runbook: High Error Rate

## Symptoms
- >1% HTTP 5xx errors over 5 minutes
- Slack alert in #monitoring

## 1. Check Traefik logs
`docker logs traefik --tail 100 | grep "5[0-9][0-9]"`

## 2. Check service logs
`docker logs <service> --tail 200`

## 3. Check database
- Connection pool (Redis: `INFO clients`)
- Query performance (`EXPLAIN ANALYZE` on slow queries)

## 4. Rollback if needed
- Revert to previous stable image: `docker compose up -d <service>@sha256:<prev>`
```

### IR Checklist
1. **Acknowledge** — confirm you're investigating.
2. **Mitigate** — stop the bleeding (rollback, disable feature flag, scale up).
3. **Resolve** — apply the fix, verify metrics return to baseline.
4. **Review** — postmortem (blameless). What happened? Why? How to prevent?

## Logging Integration

- **Structured logs** (JSON) indexed by Loki or ELK.
- **Correlate logs with metrics** — trace_id in both.
- **Error sampling** — capture 100% of errors, 1-10% of successful requests.

## This Repo's Monitoring Stack

| Component | Role |
|-----------|------|
| **Prometheus** | Metrics + alerting (Docker SD auto-discovery) |
| **Jaeger** | Tracing (OTLP receiver, all-in-one) |
| **Traefik** | Exposes metrics (--metrics.prometheus=true) |
| **Docker labels** | `prometheus.io/scrape=true` for auto-discovery |
| **Dashboard** | `/dashboard` (auto-refresh 15s) |
| **Dashboard API** | `/api/dashboard` — JSON with containers, traces, metrics |

## Anti-patterns

- ❌ No SLOs — "everything should be fast" is not a target
- ❌ Dashboard overload — metrics vomit with no narrative
- ❌ Alert fatigue — 100 alerts per day means none are actionable
- ❌ No runbooks — "what do I do when this alert fires?"
- ❌ Only monitoring infrastructure — no business metrics (orders/min, signups)
- ❌ Not monitoring after hours — 24/7 service needs 24/7 monitoring
- ❌ No log retention policy — infinite logs = infinite cost