计划执行与检查节点
executing-plans
按既定实施计划完成工作,保持过程检查和反馈。
conduct-post-mortem【效率与协作】在事故发生后开展无责备后复盘分析。构建时间线重建、识别促成因素、 并生成可操作的改进措施。专注于系统性问题而非个人责任。适用于任何生产事故或服务降级之后、 险情发生后、调查反复出现的问题,或向跨团队分享系统性经验教训。
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
主导无责备后复盘,从事故中学习并提升系统韧性。
收集事故的所有相关产物:
# Export relevant logs (adjust timerange)
kubectl logs deployment/api-service \
--since-time="2025-02-09T10:00:00Z" \
--until-time="2025-02-09T11:30:00Z" > incident-logs.txt
# Export Prometheus metrics snapshot
curl -G 'http://prometheus:9090/api/v1/query_range' \
--data-urlencode 'query=rate(http_requests_total{job="api"}[5m])' \
--data-urlencode 'start=2025-02-09T10:00:00Z' \
--data-urlencode 'end=2025-02-09T11:30:00Z' \
--data-urlencode 'step=15s' > metrics.json
# Export alert history
amtool alert query --within=2h alertname="HighErrorRate" --output json > alerts.json
预期结果: 覆盖完整事故时间线的日志、指标和告警。
失败处理: 如果数据不完整,在报告中记录缺口。为下次建立更长的保留期。
创建按时间顺序的重建:
## Timeline (all times UTC)
| Time | Event | Source | Actor |
|----------|-------|--------|-------|
| 10:05:23 | First 5xx errors appear | nginx access logs | - |
| 10:06:45 | High error rate alert fires | Prometheus | - |
| 10:08:12 | On-call engineer paged | PagerDuty | System |
| 10:12:00 | Engineer acknowledges alert | PagerDuty | @alice |
| 10:15:30 | Database connection pool exhausted | app logs | - |
| 10:18:45 | Database queries identified as slow | pganalyze | @alice |
| 10:22:10 | Cache layer deployed as mitigation | kubectl | @alice |
| 10:35:00 | Error rate returns to normal | Prometheus | - |
| 10:40:00 | Incident marked resolved | PagerDuty | @alice |
预期结果: 清晰的分钟级序列,展示发生了什么以及何时发生。
失败处理: 时间戳不匹配。确保所有系统使用 NTP 并以 UTC 记录日志。
使用五个为什么或鱼骨分析:
## Contributing Factors
### Immediate Cause
- Database connection pool exhausted (max 20 connections)
- Query introduced in v2.3.0 deployment lacked index
### Contributing Factors
1. **Monitoring Gap**: Connection pool utilization not monitored
2. **Testing Gap**: Load testing didn't include new query pattern
3. **Runbook Gap**: No documented procedure for DB connection issues
4. **Capacity Planning**: Pool size unchanged despite 3x traffic growth
### Systemic Issues
- No pre-deployment query plan review
- Database alerts only fire on total failure, not degradation
预期结果: 识别出多层次因果关系,避免指责。
失败处理: 如果分析停留在"工程师犯了错误",继续深挖。是什么允许了这个错误发生?
创建具体且可追踪的改进措施:
## Action Items
| ID | Action | Owner | Deadline | Priority |
|----|--------|-------|----------|----------|
| AI-001 | Add connection pool metrics to Grafana | @bob | 2025-02-16 | High |
| AI-002 | Create runbook: DB connection saturation | @alice | 2025-02-20 | High |
| AI-003 | Add DB query plan check to CI/CD | @charlie | 2025-03-01 | Medium |
| AI-004 | Review and adjust connection pool size | @dan | 2025-02-14 | High |
| AI-005 | Implement DB slow query alerts (<100ms) | @bob | 2025-02-23 | Medium |
| AI-006 | Add load testing for new query patterns | @charlie | 2025-03-15 | Low |
预期结果: 每个行动项都有负责人、截止日期和明确的交付物。
失败处理: "改善测试"之类的模糊行动项不会被执行。要具体明确。
使用如下模板结构:
# Post-Mortem: API Service Degradation (2025-02-09)
**Date**: 2025-02-09
**Duration**: 1h 35min (10:05 - 11:40 UTC)
**Severity**: P1 (Critical service degraded)
**Authors**: @alice, @bob
**Reviewed**: 2025-02-10
## Summary
The API service experienced elevated error rates (40% of requests) due to
database connection pool exhaustion. Service was restored by deploying a
cache layer. No data loss occurred.
## Impact
- 40,000 failed requests over 1.5 hours
- 2,000 customers affected
- Revenue impact: ~$5,000 (estimated)
## Root Cause
Query introduced in v2.3.0 deployment performed a full table scan due to
missing index. Under increased load, this saturated the connection pool.
[... timeline, contributing factors, action items as above ...]
## What Went Well
- Alert fired within 90 seconds of first errors
- Mitigation deployed quickly (10 minutes from page to fix)
- Communication to customers was clear and timely
## Lessons Learned
- Database monitoring is insufficient; need connection-level metrics
- Load testing must cover new query patterns, not just volume
- Connection pool sizing hasn't kept pace with traffic growth
## Prevention
See Action Items above.
预期结果: 报告在事故发生后 48 小时内与团队和利益相关方共享。
失败处理: 如果报告延迟超过 1 周,洞察会变得陈旧。优先处理后复盘。
追踪行动项进度:
# Create GitHub issues from action items
gh issue create --title "AI-001: Add connection pool metrics" \
--body "From post-mortem PM-2025-02-09. Owner: @bob. Deadline: 2025-02-16" \
--label "post-mortem,observability" \
--assignee bob
# Set up recurring reminder
# Add to team calendar: Weekly review of open post-mortem items
预期结果: 行动项在项目管理工具中追踪,每周审查。
失败处理: 如果行动项长期搁置,事故将重演。为高优先级事项指定高管赞助人。
write-incident-runbook - 创建事故期间引用的运行手册configure-alerting-rules - 根据后复盘发现改善告警