Conduct Post-Mortem

Lead blameless post-mortem to learn from incidents, improve system resilience.

When Use

• After any production incident or service degradation
• Following near-miss or close call
• Investigating recurring issues
• Share learnings across teams

Inputs

• Required: Incident details (start/end time, services affected, severity)
• Required: Access to logs, metrics, alerts during incident window
• Optional: Runbook used during incident response
• Optional: Communication logs (Slack, PagerDuty)

Steps

Step 1: Collect Raw Data

Gather all artifacts from incident:

# 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

Got: Logs, metrics, alerts covering full incident timeline.

If fail: Data incomplete? Note gaps in report. Set up longer retention for next time.

Step 2: Build Timeline

Create chronological reconstruction:

## 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 |

Got: Clear minute-by-minute sequence showing what happened and when.

If fail: Timestamp mismatches? Ensure all systems use NTP, log in UTC.

Step 3: Identify Contributing Factors

Use Five Whys or fishbone analysis:

## 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

Got: Multiple layers of causation identified, avoiding blame.

If fail: Analysis stops at "engineer made mistake"? Dig deeper. What allowed that mistake?

Step 4: Generate Action Items

Create concrete, trackable improvements:

## 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 |

Got: Each action has owner, deadline, clear deliverable.

If fail: Vague actions like "improve testing" won't get done. Make specific.

Step 5: Write and Distribute Report

Use this template structure:

# 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.

Got: Report shared with team, stakeholders within 48 hours of incident.

If fail: Report delays exceed 1 week? Insights grow stale. Prioritize post-mortems.

Step 6: Review Action Items in Standup/Retros

Track action item progress:

# 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

Got: Action items tracked in project management tool, reviewed weekly.

If fail: Action items languish? Incidents will recur. Assign executive sponsor for high-priority items.

Checks

• Timeline complete and chronologically accurate
• Multiple contributing factors identified (not just one)
• Action items have owners, deadlines, priorities
• Report uses blameless language (no "X caused the issue")
• Report distributed to all stakeholders within 48 hours
• Action items tracked in ticketing system
• Follow-up review scheduled for 4 weeks out

Pitfalls

• Blame culture: Using "who" language instead of "what/why". Focus on systems, not people.
• Shallow analysis: Stopping at first cause. Always ask "why" at least 5 times.
• Vague action items: "Improve monitoring" not actionable. "Add metric X to dashboard Y by date Z" is.
• No follow-through: Action items created but never reviewed. Set calendar reminders.
• Fear of transparency: Hiding incidents reduces learning. Share widely (within appropriate security boundaries).

See Also

• write-incident-runbook - create runbooks referenced during incidents
• configure-alerting-rules - improve alerts based on post-mortem findings