managing-incidents（管理事件）

Incident Management

Provide end-to-end incident management guidance covering detection, response, communication, and learning. Emphasizes SRE culture, blameless post-mortems, and structured processes for high-reliability operations.

When to Use This Skill

Apply this skill when:

• Setting up incident response processes for a team
• Designing on-call rotations and escalation policies
• Creating runbooks for common failure scenarios
• Conducting blameless post-mortems after incidents
• Implementing incident communication protocols (internal and external)
• Choosing incident management tooling and platforms
• Improving MTTR and incident frequency metrics

Core Principles

Incident Management Philosophy

Declare Early and Often: Do not wait for certainty. Declaring an incident enables coordination, can be downgraded if needed, and prevents delayed response.

Mitigation First, Root Cause Later: Stop customer impact immediately (rollback, disable feature, failover). Debug and fix root cause after stability restored.

Blameless Culture: Assume good intentions. Focus on how systems failed, not who failed. Create psychological safety for honest learning.

Clear Command Structure: Assign Incident Commander (IC) to own coordination. IC delegates tasks but does not do hands-on debugging.

Communication is Critical: Internal coordination via dedicated channels, external transparency via status pages. Update stakeholders every 15-30 minutes during critical incidents.

Severity Classification

Standard severity levels with response times:

SEV0 (P0) - Critical Outage:

• Impact: Complete service outage, critical data loss, payment processing down
• Response: Page immediately 24/7, all hands on deck, executive notification
• Example: API completely down, entire customer base affected

SEV1 (P1) - Major Degradation:

• Impact: Major functionality degraded, significant customer subset affected
• Response: Page during business hours, escalate off-hours, IC assigned
• Example: 15% error rate, critical feature unavailable

SEV2 (P2) - Minor Issues:

• Impact: Minor functionality impaired, edge case bug, small user subset
• Response: Email/Slack alert, next business day response
• Example: UI glitch, non-critical feature slow

SEV3 (P3) - Low Impact:

• Impact: Cosmetic issues, no customer functionality affected
• Response: Ticket queue, planned sprint
• Example: Visual inconsistency, documentation error

For detailed severity decision framework and interactive classifier, see references/severity-classification.md.

Incident Roles

Incident Commander (IC):

• Owns overall incident response and coordination
• Makes strategic decisions (rollback vs. debug, when to escalate)
• Delegates tasks to responders (does NOT do hands-on debugging)
• Declares incident resolved when stability confirmed

Communications Lead:

• Posts status updates to internal and external channels
• Coordinates with stakeholders (executives, product, support)
• Drafts post-incident customer communication
• Cadence: Every 15-30 minutes for SEV0/SEV1

Subject Matter Experts (SMEs):

• Hands-on debugging and mitigation
• Execute runbooks and implement fixes
• Provide technical context to IC

Scribe:

• Documents timeline, actions, decisions in real-time
• Records incident notes for post-mortem reconstruction

Assign roles based on severity:

• SEV2/SEV3: Single responder
• SEV1: IC + SME(s)
• SEV0: IC + Communications Lead + SME(s) + Scribe

For detailed role responsibilities, see references/incident-roles.md.

On-Call Management

Rotation Patterns

Primary + Secondary:

• Primary: First responder
• Secondary: Backup if primary doesn't ack within 5 minutes
• Rotation length: 1 week (optimal balance)

Follow-the-Sun (24/7):

• Team A: US hours, Team B: Europe hours, Team C: Asia hours
• Benefit: No night shifts, improved work-life balance
• Requires: Multiple global teams

Tiered Escalation:

• Tier 1: Junior on-call (common issues, runbook-driven)
• Tier 2: Senior on-call (complex troubleshooting)
• Tier 3: Team lead/architect (critical decisions)

Best Practices

• Rotation length: 1 week per rotation
• Handoff ceremony: 30-minute call to discuss active issues
• Compensation: On-call stipend + time off after major incidents
• Tooling: PagerDuty, Opsgenie, or incident.io
• Limits: Max 2-3 pages per night; escalate if exceeded

Incident Response Workflow

Standard incident lifecycle:

Detection → Triage → Declaration → Investigation
  ↓
Mitigation → Resolution → Monitoring → Closure
  ↓
Post-Mortem (within 48 hours)

Key Decision Points

When to Declare: When in doubt, declare (can always downgrade severity)

When to Escalate:

• No progress after 30 minutes
• Severity increases (SEV2 → SEV1)
• Specialized expertise needed

When to Close:

• Issue resolved and stable for 30+ minutes
• Monitoring shows all metrics at baseline
• No customer-reported issues

For complete workflow details, see references/incident-workflow.md.

Communication Protocols

Internal Communication

Incident Slack Channel:

• Format: #incident-YYYY-MM-DD-topic-description
• Pin: Severity, IC name, status update template, runbook links

War Room: Video call for SEV0/SEV1 requiring real-time voice coordination

Status Update Cadence:

• SEV0: Every 15 minutes
• SEV1: Every 30 minutes
• SEV2: Every 1-2 hours or at major milestones

External Communication

Status Page:

• Tools: Statuspage.io, Instatus, custom
• Stages: Investigating → Identified → Monitoring → Resolved
• Transparency: Acknowledge issue publicly, provide ETAs when possible

Customer Email:

• When: SEV0/SEV1 affecting customers
• Timing: Within 1 hour (acknowledge), post-resolution (full details)
• Tone: Apologetic, transparent, action-oriented

Regulatory Notifications:

• Data Breach: GDPR requires notification within 72 hours
• Financial Services: Immediate notification to regulators
• Healthcare: HIPAA breach notification rules

For communication templates, see examples/communication-templates.md.

Runbooks and Playbooks

Runbook Structure

Every runbook should include:

1. Trigger: Alert conditions that activate this runbook
2. Severity: Expected severity level
3. Prerequisites: System state requirements
4. Steps: Numbered, executable commands (copy-pasteable)
5. Verification: How to confirm fix worked
6. Rollback: How to undo if steps fail
7. Owner: Team/person responsible
8. Last Updated: Date of last revision

Best Practices

• Executable: Commands copy-pasteable, not just descriptions
• Tested: Run during disaster recovery drills
• Versioned: Track changes in Git
• Linked: Reference from alert definitions
• Automated: Convert manual steps to scripts over time

For runbook templates, see examples/runbooks/ directory.

Blameless Post-Mortems

Blameless Culture Tenets

Assume Good Intentions: Everyone made the best decision with information available.

Focus on Systems: Investigate how processes failed, not who failed.

Psychological Safety: Create environment where honesty is rewarded.

Learning Opportunity: Incidents are gifts of organizational knowledge.

Post-Mortem Process

1. Schedule Review (Within 48 Hours): While memory is fresh

2. Pre-Work: Reconstruct timeline, gather metrics/logs, draft document

3. Meeting Facilitation:

• Timeline walkthrough
• 5 Whys Analysis to identify systemic root causes
• What Went Well / What Went Wrong
• Define action items with owners and due dates

4. Post-Mortem Document:

• Sections: Summary, Timeline, Root Cause, Impact, What Went Well/Wrong, Action Items
• Distribution: Engineering, product, support, leadership
• Storage: Archive in searchable knowledge base

5. Follow-Up: Track action items in sprint planning

For detailed facilitation guide and template, see references/blameless-postmortems.md and examples/postmortem-template.md.

Alert Design Principles

Actionable Alerts Only:

• Every alert requires human action
• Include graphs, runbook links, recent changes
• Deduplicate related alerts
• Route to appropriate team based on service ownership

Preventing Alert Fatigue:

• Audit alerts quarterly: Remove non-actionable alerts
• Increase thresholds for noisy metrics
• Use anomaly detection instead of static thresholds
• Limit: Max 2-3 pages per night

Tool Selection

Incident Management Platforms

PagerDuty:

• Best for: Established enterprises, complex escalation policies
• Cost: $19-41/user/month
• When: Team size 10+, budget $500+/month

Opsgenie:

• Best for: Atlassian ecosystem users, flexible routing
• Cost: $9-29/user/month
• When: Using Atlassian products, budget $200-500/month

incident.io:

• Best for: Modern teams, AI-powered response, Slack-native
• When: Team size 5-50, Slack-centric culture

For detailed tool comparison, see references/tool-comparison.md.

Status Page Solutions

Statuspage.io: Most trusted, easy setup ($29-399/month) Instatus: Budget-friendly, modern design ($19-99/month)

Metrics and Continuous Improvement

Key Incident Metrics

MTTA (Mean Time To Acknowledge):

• Target: < 5 minutes for SEV1
• Improvement: Better on-call coverage

MTTR (Mean Time To Recovery):

• Target: < 1 hour for SEV1
• Improvement: Runbooks, automation

MTBF (Mean Time Between Failures):

• Target: > 30 days for critical services
• Improvement: Root cause fixes

Incident Frequency:

• Track: SEV0, SEV1, SEV2 counts per month
• Target: Downward trend

Action Item Completion Rate:

• Target: > 90%
• Improvement: Sprint integration, ownership clarity

Continuous Improvement Loop

Incident → Post-Mortem → Action Items → Prevention
   ↑                                          ↓
   └──────────── Fewer Incidents ─────────────┘

Decision Frameworks

Severity Classification Decision Tree

Is production completely down or critical data at risk?
├─ YES → SEV0
└─ NO  → Is major functionality degraded?
          ├─ YES → Is there a workaround?
          │        ├─ YES → SEV1
          │        └─ NO  → SEV0
          └─ NO  → Are customers impacted?
                   ├─ YES → SEV2
                   └─ NO  → SEV3

Use interactive classifier: python scripts/classify-severity.py

Escalation Matrix

For detailed escalation guidance, see references/escalation-matrix.md.

Mitigation vs. Root Cause

Prioritize Mitigation When:

• Active customer impact ongoing
• Quick fix available (rollback, disable feature)

Prioritize Root Cause When:

• Customer impact already mitigated
• Fix requires careful analysis

Default: Mitigation first (99% of cases)

Anti-Patterns to Avoid

• Delayed Declaration: Waiting for certainty before declaring incident
• Skipping Post-Mortems: "Small" incidents still provide learning
• Blame Culture: Punishing individuals prevents systemic learning
• Ignoring Action Items: Post-mortems without follow-through waste time
• No Clear IC: Multiple people leading creates confusion
• Alert Fatigue: Noisy, non-actionable alerts cause on-call to ignore pages
• Hands-On IC: IC should delegate debugging, not do it themselves

Implementation Checklist

Phase 1: Foundation (Week 1)

• Define severity levels (SEV0-SEV3)
• Choose incident management platform
• Set up basic on-call rotation
• Create incident Slack channel template

Phase 2: Processes (Weeks 2-3)

• Create first 5 runbooks for common incidents
• Set up status page
• Train team on incident response
• Conduct tabletop exercise

Phase 3: Culture (Weeks 4+)

• Conduct first blameless post-mortem
• Establish post-mortem cadence
• Implement MTTA/MTTR dashboards
• Track action items in sprint planning

Phase 4: Optimization (Months 3-6)

• Automate incident declaration
• Implement runbook automation
• Monthly disaster recovery drills
• Quarterly incident trend reviews

Integration with Other Skills

Observability: Monitoring alerts trigger incidents → Use incident-management for response

Disaster Recovery: DR provides recovery procedures → Incident-management provides operational response

Security Incident Response: Similar process with added compliance/forensics

Infrastructure-as-Code: IaC enables fast recovery via automated rebuild

Performance Engineering: Performance incidents trigger response → Performance team investigates post-mitigation

Examples and Templates

Runbook Templates:

• examples/runbooks/database-failover.md
• examples/runbooks/cache-invalidation.md
• examples/runbooks/ddos-mitigation.md

Post-Mortem Template:

• examples/postmortem-template.md - Complete blameless post-mortem structure

Communication Templates:

• examples/communication-templates.md - Status updates, customer emails

On-Call Handoff:

• examples/oncall-handoff-template.md - Weekly handoff format

Integration Scripts:

• examples/integrations/pagerduty-slack.py
• examples/integrations/statuspage-auto-update.py
• examples/integrations/postmortem-generator.py

Scripts

Interactive Severity Classifier:

python scripts/classify-severity.py

Asks questions to determine appropriate severity level based on impact and urgency.

Further Reading

Books:

• Google SRE Book: "Postmortem Culture" (Chapter 15)
• "The Phoenix Project" by Gene Kim
• "Site Reliability Engineering" (Full book)

Online Resources:

• Atlassian: "How to Run a Blameless Postmortem"
• PagerDuty: "Incident Response Guide"
• Google SRE: "Postmortem Culture: Learning from Failure"

Standards:

• Incident Command System (ICS) - FEMA standard adapted for tech
• ITIL Incident Management - Traditional IT service management