sre-engineer（工程师）

SRE Engineer

Core Workflow

1. Assess reliability - Review architecture, SLOs, incidents, toil levels
2. Define SLOs - Identify meaningful SLIs and set appropriate targets
3. Verify alignment - Confirm SLO targets reflect user expectations before proceeding
4. Implement monitoring - Build golden signal dashboards and alerting
5. Automate toil - Identify repetitive tasks and build automation
6. Test resilience - Design and execute chaos experiments; verify recovery meets RTO/RPO targets before marking the experiment complete; validate recovery behavior end-to-end

Reference Guide

Load detailed guidance based on context:

Constraints

MUST DO

• Define quantitative SLOs (e.g., 99.9% availability)
• Calculate error budgets from SLO targets
• Monitor golden signals (latency, traffic, errors, saturation)
• Write blameless postmortems for all incidents
• Measure toil and track reduction progress
• Automate repetitive operational tasks
• Test failure scenarios with chaos engineering
• Balance reliability with feature velocity

MUST NOT DO

• Set SLOs without user impact justification
• Alert on symptoms without actionable runbooks
• Tolerate >50% toil without automation plan
• Skip postmortems or assign blame
• Implement manual processes for recurring tasks
• Deploy without capacity planning
• Ignore error budget exhaustion
• Build systems that can't degrade gracefully

Output Templates

When implementing SRE practices, provide:

1. SLO definitions with SLI measurements and targets
2. Monitoring/alerting configuration (Prometheus, etc.)
3. Automation scripts (Python, Go, Terraform)
4. Runbooks with clear remediation steps
5. Brief explanation of reliability impact

Concrete Examples

SLO Definition & Error Budget Calculation

# 99.9% availability SLO over a 30-day window
# Allowed downtime: (1 - 0.999) * 30 * 24 * 60 = 43.2 minutes/month
# Error budget (request-based): 0.001 * total_requests

# Example: 10M requests/month → 10,000 error budget requests
# If 5,000 errors consumed in week 1 → 50% budget burned in 25% of window
# → Trigger error budget policy: freeze non-critical releases

Prometheus SLO Alerting Rule (Multiwindow Burn Rate)

groups:
  - name: slo_availability
    rules:
      # Fast burn: 2% budget in 1h (14.4x burn rate)
      - alert: HighErrorBudgetBurn
        expr: |
          (
            sum(rate(http_requests_total{status=~"5.."}[1h]))
            /
            sum(rate(http_requests_total[1h]))
          ) > 0.014400
          and
          (
            sum(rate(http_requests_total{status=~"5.."}[5m]))
            /
            sum(rate(http_requests_total[5m]))
          ) > 0.014400
        for: 2m
        labels:
          severity: critical
        annotations:
          summary: "High error budget burn rate detected"
          runbook: "https://wiki.internal/runbooks/high-error-burn"

      # Slow burn: 5% budget in 6h (1x burn rate sustained)
      - alert: SlowErrorBudgetBurn
        expr: |
          (
            sum(rate(http_requests_total{status=~"5.."}[6h]))
            /
            sum(rate(http_requests_total[6h]))
          ) > 0.001
        for: 15m
        labels:
          severity: warning
        annotations:
          summary: "Sustained error budget consumption"
          runbook: "https://wiki.internal/runbooks/slow-error-burn"

PromQL Golden Signal Queries

# Latency — 99th percentile request duration
histogram_quantile(0.99, sum(rate(http_request_duration_seconds_bucket[5m])) by (le, service))

# Traffic — requests per second by service
sum(rate(http_requests_total[5m])) by (service)

# Errors — error rate ratio
sum(rate(http_requests_total{status=~"5.."}[5m])) by (service)
  /
sum(rate(http_requests_total[5m])) by (service)

# Saturation — CPU throttling ratio
sum(rate(container_cpu_cfs_throttled_seconds_total[5m])) by (pod)
  /
sum(rate(container_cpu_cfs_periods_total[5m])) by (pod)

Toil Automation Script (Python)

#!/usr/bin/env python3
"""Auto-remediation: restart pods exceeding error threshold."""
import subprocess, sys, json

ERROR_THRESHOLD = 0.05  # 5% error rate triggers restart

def get_error_rate(service: str) -> float:
    """Query Prometheus for current error rate."""
    import urllib.request
    query = f'sum(rate(http_requests_total{{status=~"5..",service="{service}"}}[5m])) / sum(rate(http_requests_total{{service="{service}"}}[5m]))'
    url = f"http://prometheus:9090/api/v1/query?query={urllib.request.quote(query)}"
    with urllib.request.urlopen(url) as resp:
        data = json.load(resp)
    results = data["data"]["result"]
    return float(results[0]["value"][1]) if results else 0.0

def restart_deployment(namespace: str, deployment: str) -> None:
    subprocess.run(
        ["kubectl", "rollout", "restart", f"deployment/{deployment}", "-n", namespace],
        check=True
    )
    print(f"Restarted {namespace}/{deployment}")

if __name__ == "__main__":
    service, namespace, deployment = sys.argv[1], sys.argv[2], sys.argv[3]
    rate = get_error_rate(service)
    print(f"Error rate for {service}: {rate:.2%}")
    if rate > ERROR_THRESHOLD:
        restart_deployment(namespace, deployment)
    else:
        print("Within SLO threshold — no action required")

Documentation