技能创建器
skill-creator
创建、改进和评估自定义技能,把专业流程转化为可重复使用的 AI 能力。
define-slo-sli-sla【数据与研究】Establish Service Level Objectives (SLO), Service Level Indicators (SLI), and Service Level Agreements (SLA) with error budget tracking, burn rate alerts, and automated reporting using Prometheus and tools like Sloth or Pyrra. Use when defining reliability targets for customer-facing services, balancing feature velocity against system reliability thr…
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
以 SLO 立可測之信賴標的,以 SLI 察之,以錯額管之。
詳見 Extended Examples 全備之配置模板。
察此三者之別與所連。
義:
SLI (Service Level Indicator)
- **What**: A quantitative measure of service behavior
- **Example**: Request success rate, request latency, system throughput
- **Measurement**: `successful_requests / total_requests * 100`
SLO (Service Level Objective)
- **What**: Target value or range for an SLI over a time window
- **Example**: 99.9% of requests succeed in 30-day window
- **Purpose**: Internal reliability target to guide operations
SLA (Service Level Agreement)
- **What**: Contractual commitment with consequences for missing SLO
- **Example**: 99.9% uptime SLA with refunds if breached
- **Purpose**: External promise to customers with penalties
次第:
SLA (99.9% uptime, customer refunds)
├─ SLO (99.95% success rate, internal target)
│ └─ SLI (actual measured: 99.97% success rate)
└─ Error Budget (0.05% failures allowed per month)
要:SLO 宜嚴於 SLA,以作客受影之前之緩衝。
舉例:
得: 組員明其別,約哪些度為 SLI,同定 SLO 之標。
敗則:
擇反映用者體驗與業之衝擊者。
四黃金信號(Google SRE):
延時:服一請求所費之時
# P95 latency
histogram_quantile(0.95,
sum(rate(http_request_duration_seconds_bucket[5m])) by (le)
)
流量:系統所受之需
# Requests per second
sum(rate(http_requests_total[5m]))
錯:敗之率
# Error rate percentage
sum(rate(http_requests_total{status=~"5.."}[5m]))
/ sum(rate(http_requests_total[5m])) * 100
飽和:系統已盈幾何
# CPU saturation
avg(rate(node_cpu_seconds_total{mode!="idle"}[5m]))
常見 SLI 模式:
# Availability SLI
availability:
description: "Percentage of successful requests"
query: |
sum(rate(http_requests_total{status!~"5.."}[5m]))
/ sum(rate(http_requests_total[5m]))
good_threshold: 0.999 # 99.9%
# Latency SLI
latency:
description: "P99 request latency under 500ms"
query: |
histogram_quantile(0.99,
sum(rate(http_request_duration_seconds_bucket[5m])) by (le)
) < 0.5
good_threshold: 0.95 # 95% of windows meet target
# Throughput SLI
throughput:
description: "Requests processed per second"
query: |
sum(rate(http_requests_total[5m]))
good_threshold: 1000 # Minimum 1000 req/s
# Data freshness SLI (for batch jobs)
freshness:
description: "Data updated within last hour"
query: |
(time() - max(data_last_updated_timestamp)) < 3600
good_threshold: 1 # Always fresh
SLI 擇之則:
宜避者:
得: 每服擇 2-4 SLI,最少含可用與延時;團隊同意其度查語。
敗則:
定實際可達之信賴標。
SLO 規格式:
service: user-api
slos:
- name: availability
objective: 99.9
description: |
99.9% of requests return non-5xx status codes
# ... (see EXAMPLES.md for complete configuration)
擇時窗:
常見者:
三十日窗錯額例:
SLO: 99.9% availability over 30 days
Allowed failures: 0.1%
Total requests per month: 100M
Error budget: 100,000 failed requests
Daily budget: ~3,333 failed requests
立實際之標:
基於當前之表現:
# Check actual availability over past 90 days
avg_over_time(
(sum(rate(http_requests_total{status!~"5.."}[5m]))
/ sum(rate(http_requests_total[5m])))[90d:5m]
)
# Result: 99.95% → Set SLO at 99.9% (safer than current)
算九之價:
99% → 7.2 hours downtime/month (low reliability)
99.9% → 43 minutes downtime/month (good)
99.95% → 22 minutes downtime/month (very good)
99.99% → 4.3 minutes downtime/month (expensive)
99.999% → 26 seconds downtime/month (very expensive)
衡用者之樂與工程之費:
得: SLO 標經業關者同意而立,有由有據,錯額已算。
敗則:
以 Sloth 由 SLO 規生 Prometheus 記錄律與警。
裝 Sloth:
# Binary installation
wget https://github.com/slok/sloth/releases/download/v0.11.0/sloth-linux-amd64
chmod +x sloth-linux-amd64
sudo mv sloth-linux-amd64 /usr/local/bin/sloth
# Or Docker
docker pull ghcr.io/slok/sloth:latest
建 Sloth SLO 規(slos/user-api.yml):
version: "prometheus/v1"
service: "user-api"
labels:
team: "platform"
tier: "1"
slos:
# ... (see EXAMPLES.md for complete configuration)
生 Prometheus 律:
# Generate recording and alerting rules
sloth generate -i slos/user-api.yml -o prometheus/rules/user-api-slo.yml
# Validate generated rules
promtool check rules prometheus/rules/user-api-slo.yml
所生之記錄律(節錄):
groups:
- name: sloth-slo-sli-recordings-user-api-requests-availability
interval: 30s
rules:
# SLI: Ratio of good events
- record: slo:sli_error:ratio_rate5m
# ... (see EXAMPLES.md for complete configuration)
所生之警:
groups:
- name: sloth-slo-alerts-user-api-requests-availability
rules:
# Fast burn: 2% budget consumed in 1 hour
- alert: UserAPIHighErrorRate
expr: |
# ... (see EXAMPLES.md for complete configuration)
載律入 Prometheus:
# prometheus.yml
rule_files:
- "rules/user-api-slo.yml"
重載 Prometheus:
curl -X POST http://localhost:9090/-/reload
得: Sloth 生多窗多燒率警,記錄律成評,事故時警適發。
敗則:
yamllint slos/user-api.yml 驗 YAML 法curl http://localhost:9090/api/v1/rules於 Grafana 可視化 SLO 合規與錯額耗。
Grafana 儀表盤 JSON(節錄):
{
"dashboard": {
"title": "SLO Dashboard - User API",
"panels": [
{
"type": "stat",
# ... (see EXAMPLES.md for complete configuration)
宜視之要度:
錯額政策盤(markdown 板):
## Error Budget Policy
**Current Status**: 78% budget remaining
### If Error Budget > 50%
- ✅ Full speed ahead on new features
# ... (see EXAMPLES.md for complete configuration)
得: 儀表盤示實時 SLO 合規,錯額耗可見,團隊可明斷新功之速。
敗則:
curl http://localhost:9090/api/v1/rules | jq '.data.groups[].rules[] | select(.name | contains("slo:"))'定組織管錯額之程。
錯額政策模板:
service: user-api
slo:
availability: 99.9%
latency_p99: 200ms
window: 30 days
# ... (see EXAMPLES.md for complete configuration)
自動執政:
# Example: Deployment gate script
import requests
import sys
def check_error_budget(service):
# Query Prometheus for error budget
# ... (see EXAMPLES.md for complete configuration)
合入 CI/CD 流程:
# .github/workflows/deploy.yml
jobs:
check-error-budget:
runs-on: ubuntu-latest
steps:
- name: Check SLO Error Budget
run: |
python scripts/check_error_budget.py user-api
- name: Deploy
if: success()
run: |
kubectl apply -f deploy/
得: 政策已記明,自動閘阻額耗時之險部署,團隊同其信賴之要。
敗則:
setup-prometheus-monitoring - 配 Prometheus 以集度供 SLI 算configure-alerting-rules - 合 SLO 燒率警入 Alertmanager 供值班告build-grafana-dashboards - 視 SLO 合規與錯額耗write-incident-runbook - 於運行手冊含 SLO 衝擊以定事故應之先