on-call-handoff-patterns

模式：认知/提示词驱动 — 没有独立的实用脚本；通过 Agent 上下文使用。

值班交接模式

有效的值班交接模式，确保跨班次的连续性、上下文传递和可靠的事件响应。

何时使用此技能

• 交接值班职责
• 撰写班次交接摘要
• 记录正在进行的调查
• 制定值班轮换流程
• 提高交接质量
• 帮助新的值班工程师熟悉工作

核心概念

1. 交接内容

内容	用途
活动事件	当前出现故障的内容
正在进行的调查	正在调试的问题
近期变更	部署、配置
已知问题	正在使用的临时解决方案
即将发生的事件	维护、发布

2. 交接时间

Recommended: 30 min overlap between shifts

Outgoing:
├── 15 min: Write handoff document
└── 15 min: Sync call with incoming

Incoming:
├── 15 min: Review handoff document
├── 15 min: Sync call with outgoing
└── 5 min: Verify alerting setup

模板

模板 1：班次交接文档

# On-Call Handoff: Platform Team

**Outgoing**: @alice (2024-01-15 to 2024-01-22)
**Incoming**: @bob (2024-01-22 to 2024-01-29)
**Handoff Time**: 2024-01-22 09:00 UTC

---

## Active Incidents

### None currently active

No active incidents at handoff time.

---

## Ongoing Investigations

### 1. Intermittent API Timeouts (ENG-1234)

**Status**: Investigating
**Started**: 2024-01-20
**Impact**: ~0.1% of requests timing out

**Context**:

- Timeouts correlate with database backup window (02:00-03:00 UTC)
- Suspect backup process causing lock contention
- Added extra logging in PR #567 (deployed 01/21)

**Next Steps**:

- [ ] Review new logs after tonight's backup
- [ ] Consider moving backup window if confirmed

**Resources**:

- Dashboard: [API Latency](https://grafana/d/api-latency)
- Thread: #platform-eng (01/20, 14:32)

---

### 2. Memory Growth in Auth Service (ENG-1235)

**Status**: Monitoring
**Started**: 2024-01-18
**Impact**: None yet (proactive)

**Context**:

- Memory usage growing ~5% per day
- No memory leak found in profiling
- Suspect connection pool not releasing properly

**Next Steps**:

- [ ] Review heap dump from 01/21
- [ ] Consider restart if usage > 80%

**Resources**:

- Dashboard: [Auth Service Memory](https://grafana/d/auth-memory)
- Analysis doc: [Memory Investigation](https://docs/eng-1235)

---

## Resolved This Shift

### Payment Service Outage (2024-01-19)

- **Duration**: 23 minutes
- **Root Cause**: Database connection exhaustion
- **Resolution**: Rolled back v2.3.4, increased pool size
- **Postmortem**: [POSTMORTEM-89](https://docs/postmortem-89)
- **Follow-up tickets**: ENG-1230, ENG-1231

---

## Recent Changes

### Deployments

| Service      | Version | Time        | Notes                      |
| ------------ | ------- | ----------- | -------------------------- |
| api-gateway  | v3.2.1  | 01/21 14:00 | Bug fix for header parsing |
| user-service | v2.8.0  | 01/20 10:00 | New profile features       |
| auth-service | v4.1.2  | 01/19 16:00 | Security patch             |

### Configuration Changes

- 01/21: Increased API rate limit from 1000 to 1500 RPS
- 01/20: Updated database connection pool max from 50 to 75

### Infrastructure

- 01/20: Added 2 nodes to Kubernetes cluster
- 01/19: Upgraded Redis from 6.2 to 7.0

---

## Known Issues & Workarounds

### 1. Slow Dashboard Loading

**Issue**: Grafana dashboards slow on Monday mornings
**Workaround**: Wait 5 min after 08:00 UTC for cache warm-up
**Ticket**: OPS-456 (P3)

### 2. Flaky Integration Test

**Issue**: `test_payment_flow` fails intermittently in CI
**Workaround**: Re-run failed job (usually passes on retry)
**Ticket**: ENG-1200 (P2)

---

## Upcoming Events

| Date        | Event                | Impact              | Contact       |
| ----------- | -------------------- | ------------------- | ------------- |
| 01/23 02:00 | Database maintenance | 5 min read-only     | @dba-team     |
| 01/24 14:00 | Major release v5.0   | Monitor closely     | @release-team |
| 01/25       | Marketing campaign   | 2x traffic expected | @platform     |

---

## Escalation Reminders

| Issue Type      | First Escalation     | Second Escalation |
| --------------- | -------------------- | ----------------- |
| Payment issues  | @payments-oncall     | @payments-manager |
| Auth issues     | @auth-oncall         | @security-team    |
| Database issues | @dba-team            | @infra-manager    |
| Unknown/severe  | @engineering-manager | @vp-engineering   |

---

## Quick Reference

### Common Commands

```bash
# Check service health
kubectl get pods -A | grep -v Running

# Recent deployments
kubectl get events --sort-by='.lastTimestamp' | tail -20

# Database connections
psql -c "SELECT count(*) FROM pg_stat_activity;"

# Clear cache (emergency only)
redis-cli FLUSHDB
```

重要链接

• 运行手册
• 服务目录
• 事件 Slack
• PagerDuty

---

交接检查清单

离岗工程师

• 记录活动事件
• 记录正在进行的调查
• 列出近期变更
• 说明已知问题
• 添加即将发生的事件
• 与接岗工程师同步

接岗工程师

• 阅读本文档
• 参加同步会议
• 确认 PagerDuty 会将告警路由给你
• 确认 Slack 通知正常
• 检查 VPN 和访问权限是否正常
• 查看关键仪表板

### Template 2: Quick Handoff (Async)

```markdown
# Quick Handoff: @alice → @bob

## TL;DR
- No active incidents
- 1 investigation ongoing (API timeouts, see ENG-1234)
- Major release tomorrow (01/24) - be ready for issues

## Watch List
1. API latency around 02:00-03:00 UTC (backup window)
2. Auth service memory (restart if > 80%)

## Recent
- Deployed api-gateway v3.2.1 yesterday (stable)
- Increased rate limits to 1500 RPS

## Coming Up
- 01/23 02:00 - DB maintenance (5 min read-only)
- 01/24 14:00 - v5.0 release

## Questions?
I'll be available on Slack until 17:00 today.

模板 3：事件交接（事件处理中）

# INCIDENT HANDOFF: Payment Service Degradation

**Incident Start**: 2024-01-22 08:15 UTC
**Current Status**: Mitigating
**Severity**: SEV2

---

## Current State

- Error rate: 15% (down from 40%)
- Mitigation in progress: scaling up pods
- ETA to resolution: ~30 min

## What We Know

1. Root cause: Memory pressure on payment-service pods
2. Triggered by: Unusual traffic spike (3x normal)
3. Contributing: Inefficient query in checkout flow

## What We've Done

- Scaled payment-service from 5 → 15 pods
- Enabled rate limiting on checkout endpoint
- Disabled non-critical features

## What Needs to Happen

1. Monitor error rate - should reach <1% in ~15 min
2. If not improving, escalate to @payments-manager
3. Once stable, begin root cause investigation

## Key People

- Incident Commander: @alice (handing off)
- Comms Lead: @charlie
- Technical Lead: @bob (incoming)

## Communication

- Status page: Updated at 08:45
- Customer support: Notified
- Exec team: Aware

## Resources

- Incident channel: #inc-20240122-payment
- Dashboard: [Payment Service](https://grafana/d/payments)
- Runbook: [Payment Degradation](https://wiki/runbooks/payments)

---

**Incoming on-call (@bob) - Please confirm you have:**

- [ ] Joined #inc-20240122-payment
- [ ] Access to dashboards
- [ ] Understand current state
- [ ] Know escalation path

交接同步会议

议程（15 分钟）

## Handoff Sync: @alice → @bob

1. **Active Issues** (5 min)
   - Walk through any ongoing incidents
   - Discuss investigation status
   - Transfer context and theories

2. **Recent Changes** (3 min)
   - Deployments to watch
   - Config changes
   - Known regressions

3. **Upcoming Events** (3 min)
   - Maintenance windows
   - Expected traffic changes
   - Releases planned

4. **Questions** (4 min)
   - Clarify anything unclear
   - Confirm access and alerting
   - Exchange contact info

值班最佳实践

值班前

## Pre-Shift Checklist

### Access Verification

- [ ] VPN working
- [ ] kubectl access to all clusters
- [ ] Database read access
- [ ] Log aggregator access (Splunk/Datadog)
- [ ] PagerDuty app installed and logged in

### Alerting Setup

- [ ] PagerDuty schedule shows you as primary
- [ ] Phone notifications enabled
- [ ] Slack notifications for incident channels
- [ ] Test alert received and acknowledged

### Knowledge Refresh

- [ ] Review recent incidents (past 2 weeks)
- [ ] Check service changelog
- [ ] Skim critical runbooks
- [ ] Know escalation contacts

### Environment Ready

- [ ] Laptop charged and accessible
- [ ] Phone charged
- [ ] Quiet space available for calls
- [ ] Secondary contact identified (if traveling)

值班期间

## Daily On-Call Routine

### Morning (start of day)

- [ ] Check overnight alerts
- [ ] Review dashboards for anomalies
- [ ] Check for any P0/P1 tickets created
- [ ] Skim incident channels for context

### Throughout Day

- [ ] Respond to alerts within SLA
- [ ] Document investigation progress
- [ ] Update team on significant issues
- [ ] Triage incoming pages

### End of Day

- [ ] Hand off any active issues
- [ ] Update investigation docs
- [ ] Note anything for next shift

值班后

## Post-Shift Checklist

- [ ] Complete handoff document
- [ ] Sync with incoming on-call
- [ ] Verify PagerDuty routing changed
- [ ] Close/update investigation tickets
- [ ] File postmortems for any incidents
- [ ] Take time off if shift was stressful

升级指南

何时升级

## Escalation Triggers

### Immediate Escalation

- SEV1 incident declared
- Data breach suspected
- Unable to diagnose within 30 min
- Customer or legal escalation received

### Consider Escalation

- Issue spans multiple teams
- Requires expertise you don't have
- Business impact exceeds threshold
- You're uncertain about next steps

### How to Escalate

1. Page the appropriate escalation path
2. Provide brief context in Slack
3. Stay engaged until escalation acknowledges
4. Hand off cleanly, don't just disappear

最佳实践

应做事项

• 记录所有内容 — 未来的你会感谢自己
• 尽早升级 — 宁可谨慎，也不要冒险
• 适当休息 — 告警疲劳确实存在
• 同步交接 — 异步沟通会丢失上下文
• 测试你的配置 — 在事件发生前测试，而不是发生时

不应做的事项

• 不要跳过交接 — 上下文丢失会导致事件
• 不要逞英雄 — 需要时就升级
• 不要忽略告警 — 即使看起来不严重
• 生病时不要工作 — 改换班次即可
• 不要失联 — 值班期间保持可联系

资源

• Google SRE - 值班
• PagerDuty 值班指南
• Increment 值班专题

铁律

1. 始终在结束班次前撰写交接文档 — 仅靠同步会议并不足够；口头上下文几小时内就会被遗忘；这份文档是持久记录，接岗工程师凌晨 3 点也能查阅。
2. 绝不跳过值班交接同步会议 — 仅异步交接会丢失细节、语气，以及回答问题的机会；15 分钟的重叠交接能避免接岗工程师花数小时困惑不解。
3. 遇到事件处理受阻时，始终在 30 分钟内升级 — 独自调试数小时会延误解决并耗尽值班工程师的精力；升级体现的是良好判断力，而不是软弱。
4. 绝不跳过班次开始前的告警配置验证 — 在事件处理中才发现 PagerDuty 路由配置错误就太晚了；每次值班前都要测试完整的告警链路。
5. 始终为每项未解决的调查记录明确的后续步骤 — 没有后续步骤就交接，会迫使接岗工程师从头开始；书面行动项可将交接后的熟悉时间从数小时缩短到几分钟。

反模式

反模式	失败原因	正确做法
仅异步交接（没有同步通话）	上下文丢失会导致事故；细节被遗漏；问题得不到解答	始终进行同步通话（至少重叠 15 分钟），并提供书面文档
没有交接文档	新接班的工程师不了解正在发生的问题；凌晨 3 点开始处理根本来不及	每次值班结束前都撰写完整的交接文档
值班前跳过告警验证	事故发生时才发现路由配置错误——为时已晚，无法修复	每次值班开始前测试 PagerDuty 路由和 Slack 通知
英雄主义文化（不升级问题）	个人精疲力竭；解决延迟；疲劳之下做出糟糕决策	遇到阻碍后 30 分钟内升级问题；升级是明智之举
交接调查事项时不说明后续步骤	接班工程师必须从头重新开始调查	为每个未解决事项记录当前状态、推测、证据和明确的后续步骤

记忆协议（强制）

开始前： 阅读 .claude/context/memory/learnings.md

完成后：

• 新模式 -> .claude/context/memory/learnings.md
• 发现问题 -> .claude/context/memory/issues.md
• 做出决策 -> .claude/context/memory/decisions.md

假设工作会被中断：如果没有记录在记忆中，就等于没有发生。