agentleFS
pfangueiropfangueiro/claude-code-agents/.claude/skills/sre-runbooks/SKILL.md

sre-runbooks

Skill7 starsChanged 23 days agoMITView on GitHub

SRE operational runbooks, on-call handbook, blameless postmortem templates, SLI/SLO/error budget definitions, toil reduction framework, capacity planning, and chaos engineering playbook. Auto-activates on runbook, on-call, postmortem, error budget, SLO, SLA, capacity planning, chaos engineering, incident review, toil, game day.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every P1/P2 gets a postmortem within 5 business days

    line 358

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

SRE Runbooks

Overview

Operational runbook templates and SRE practices for production reliability. Covers on-call workflows, incident response, postmortems, capacity planning, and chaos engineering.

SLI/SLO Definition Template

service: payment-api
owner: payments-team
tier: critical

slis:
  availability:
    description: "Ratio of successful HTTP responses (non-5xx) to total responses"
    query: "sum(rate(http_requests_total{service='payment-api', status!~'5..'}[5m])) / sum(rate(http_requests_total{service='payment-api'}[5m]))"
    good_event: "HTTP response with status < 500"
    valid_event: "All HTTP responses (excluding health checks)"

  latency:
    description: "Ratio of requests served faster than 300ms"
    query: "sum(rate(http_request_duration_seconds_bucket{service='payment-api', le='0.3'}[5m])) / sum(rate(http_request_duration_seconds_count{service='payment-api'}[5m]))"
    threshold: 300ms
    percentile: p99

slos:
  availability:
    target: 99.95%
    window: 30d
    error_budget: 21.9 minutes/month

  latency:
    target: 99.0%
    window: 30d
    description: "99% of requests complete within 300ms"

error_budget_policy:
  budget_available:  # >50% remaining
    - "Ship features normally"
    - "Conduct chaos experiments"
    - "Allow risky deployments with rollback plan"
  budget_warning:    # 20-50% remaining
    - "Prioritize reliability work alongside features"
    - "Review recent incidents for patterns"
    - "No chaos experiments without team approval"
  budget_critical:   # <20% remaining
    - "Freeze non-critical deployments while the root cause of the burn is unexplained or recurring"
    - "Mandatory postmortem for any new incident"
    - "Dedicated reliability sprint next cycle"
  budget_exhausted:  # 0% remaining — everything in budget_critical, plus:
    - "Escalate to engineering leadership"
    - "All engineering effort on reliability until the burn is explained"
    - "Root cause understood and mitigated: the freeze lifts and releases resume, even at zero budget"

The freeze trigger is an unexplained root cause, not budget == 0. Google's CRE team is explicit that "blocking releases solely because the 30-day error budget is exhausted" works against the policy's deliberate bias toward development velocity (Applying the escalation policy). Their escalation ladder pauses feature releases when the cause continues to evade understanding, and holds that pause until the cause is understood or the incident ages out of the 30-day SLO window. A one-off burn that is understood and mitigated brings the service back into policy without a freeze; a burn that recurs — or that nobody can explain — is what escalates.

On-Call Handbook

Rotation Structure

Primary:    First responder, handles all pages
Secondary:  Backup if primary doesn't respond in 15 minutes
Escalation: Engineering manager → VP Engineering (for P1 only)
Rotation:   Weekly, handoff every Monday 10:00 AM

On-Call Expectations

AspectRequirement
Acknowledge pageWithin 5 minutes
Begin triageWithin 15 minutes
Update status pageWithin 20 minutes (P1/P2)
Escalate if stuckOn the Escalation Matrix clock below — P1: eng manager at 15 min, VP at 30 min; P2: eng manager at 30 min; P3: team lead at 48h. The matrix is the only escalation clock
Max incidents/shift2 per 12-hour shift (exceeding = process problem)
CompensationOn-call stipend + per-incident for off-hours

Handoff Template

## On-Call Handoff: [date]

### Active Issues
- [Issue]: [status], [next action], [ETA]

### Recent Changes
- [deployment/config change]: [when], [who], [rollback plan]

### Upcoming Risks
- [planned maintenance]: [when], [impact], [owner]
- [known issue]: [workaround], [fix ETA]

### Notes for Next Shift
- [anything unusual observed]

Escalation Matrix

SeverityDescriptionResponseEscalation
P1Service fully down, data loss riskPage immediatelyEng manager at 15min, VP at 30min
P2Degraded service, partial impactPage within 15minEng manager at 30min
P3Minor impact, workaround availableTicket, next business dayTeam lead at 48h
P4Cosmetic, no user impactBacklogN/A

Infrastructure Runbooks

High CPU on Nodes

Symptoms: Node CPU >85%, pod throttling, slow responses
Alert: NodeCPUHigh

Diagnosis:
1. Identify top CPU consumers:
   kubectl top pods --all-namespaces --sort-by=cpu | head -20

2. Check if it's one pod or many:
   - One pod: check for CPU-intensive operation, missing limits, infinite loop
   - Many pods: check for traffic spike, external dependency timeout causing retries

3. Check HPA status:
   kubectl get hpa --all-namespaces

Mitigation:
- Immediate: Scale out nodes (Karpenter will auto-provision if configured)
- Short-term: Increase CPU limits for the affected pod, add HPA if missing
- Long-term: Profile the application, optimize hot paths, add caching

Verification:
- kubectl top nodes → CPU <70%
- No pods in Throttling state
- Response latency returns to normal

Pod CrashLoopBackOff

Symptoms: Pod restarts repeatedly, service degraded
Alert: PodCrashLooping

Diagnosis:
1. Check pod events:
   kubectl describe pod POD -n NAMESPACE | tail -30

2. Check previous container logs:
   kubectl logs POD -n NAMESPACE --previous

3. Common causes:
   - OOMKilled → check: kubectl get pod POD -o jsonpath='{.status.containerStatuses[0].lastState}'
   - Application error → check logs for stack trace
   - Missing config/secret → check: kubectl get pod POD -o yaml | grep -A5 envFrom
   - Failed liveness/startup probe → kubelet killed the container; compare probe
     config vs actual startup time (readiness failure cannot restart a container —
     it only pulls the Pod IP out of the Service's EndpointSlices)
   - Permission denied → check securityContext, serviceAccount

Mitigation:
- OOMKilled: Increase memory limits (check actual usage first)
- App error: Fix code, rollback deployment if recent change
- Missing config: Create/fix ConfigMap or Secret
- Liveness/startup probe failure: Increase initialDelaySeconds or add a startupProbe,
  adjust failureThreshold
- Permission: Fix RBAC, adjust securityContext

Verification:
- kubectl get pod POD → Running, 0 restarts in last 10 min
- kubectl logs POD → no error messages
- Service responding to health checks

Database Connection Pool Exhaustion

Symptoms: "too many connections", application timeouts, 503 errors
Alert: DatabaseConnectionPoolExhausted

Diagnosis:
1. Check active connections:
   SELECT count(*), state FROM pg_stat_activity GROUP BY state;

2. Check for idle-in-transaction:
   SELECT pid, query, state, query_start FROM pg_stat_activity
   WHERE state = 'idle in transaction' AND query_start < NOW() - INTERVAL '5 minutes';

3. Check for long-running queries:
   SELECT pid, query, state, NOW() - query_start AS duration
   FROM pg_stat_activity WHERE state = 'active' ORDER BY duration DESC LIMIT 10;

4. Check application pool settings vs database max_connections

Mitigation:
- Immediate: Kill idle-in-transaction connections:
  SELECT pg_terminate_backend(pid) FROM pg_stat_activity
  WHERE state = 'idle in transaction' AND query_start < NOW() - INTERVAL '10 minutes';
- Short-term: Increase pool size or max_connections (with RAM check)
- Long-term: Fix connection leaks in application, add connection timeouts, use PgBouncer

Verification:
- Active connections < 80% of max_connections
- No idle-in-transaction connections older than 1 minute
- Application 503 errors resolved

Certificate Expiration

Symptoms: TLS errors, browser warnings, API connection failures
Alert: CertificateExpiringSoon (fires at 30 days, 7 days, 1 day)

Diagnosis:
1. Check certificate expiry:
   echo | openssl s_client -connect HOST:443 2>/dev/null | openssl x509 -noout -dates

2. Check Kubernetes TLS secrets:
   kubectl get secret TLS_SECRET -o jsonpath='{.data.tls\.crt}' | base64 -d | openssl x509 -noout -dates

3. Check cert-manager status (if used):
   kubectl get certificates --all-namespaces
   kubectl describe certificate CERT_NAME -n NAMESPACE

Mitigation:
- cert-manager: Check Certificate resource, Issuer status, fix any errors
- Manual: Renew certificate, update Kubernetes secret, restart ingress controller
- ACM (AWS): Certificates auto-renew — check DNS validation records exist

Prevention:
- Use cert-manager with Let's Encrypt for automatic renewal
- Alert at 30 days, 7 days, 1 day before expiry
- Never use self-signed certificates in production

Node NotReady in Kubernetes

Symptoms: Pods evicted, scheduling failures, workload disruption
Alert: KubernetesNodeNotReady

Diagnosis:
1. Check node status:
   kubectl get nodes
   kubectl describe node NODE_NAME | grep -A10 Conditions

2. Check kubelet:
   ssh NODE → journalctl -u kubelet --since "30 minutes ago" | tail -50

3. Common causes:
   - Disk pressure: df -h (>85% = kubelet marks NotReady)
   - Memory pressure: free -m (OOM killer active)
   - Network: can node reach API server? curl -k https://APISERVER:6443/healthz
   - kubelet crash: systemctl status kubelet

Mitigation:
- Disk pressure: Clean up logs, images: crictl rmi --prune
- Memory: Identify OOM-killed processes, restart heavy workloads
- Network: Check security groups, route tables, VPN
- kubelet: Restart kubelet, check certificates

- If node won't recover:
  kubectl drain NODE_NAME --ignore-daemonsets --delete-emptydir-data
  (Karpenter will replace the node automatically)

Verification:
- kubectl get nodes → all Ready
- Pods rescheduled and running on healthy nodes
- No PodDisruptionBudget violations

Blameless Postmortem Template

# Postmortem: [Incident Title]

**Date**: YYYY-MM-DD  **Severity**: P1/P2/P3  **Duration**: X hours Y minutes
**Authors**: [names]  **Reviewers**: [names]  **Status**: Draft / Action Items In Progress / Closed

## Executive Summary
[2-3 sentences: what happened, impact, resolution]

## Impact
- **Users affected**: [count or percentage]
- **Duration**: [minutes]
- **Revenue impact**: [estimate or "not applicable"]
- **SLO impact**: [X% of error budget consumed]
- **Data loss**: [yes/no, details if yes]

## Timeline (all times in UTC)

| Time | Event |
|------|-------|
| HH:MM | Triggering event (deployment, config change, traffic spike) |
| HH:MM | First alert fired: [alert name] |
| HH:MM | On-call acknowledged |
| HH:MM | Initial triage: [what was checked] |
| HH:MM | Root cause identified |
| HH:MM | Mitigation applied: [what was done] |
| HH:MM | Service partially recovered |
| HH:MM | Service fully recovered |
| HH:MM | All-clear communicated |

## Root Cause Analysis

### 5 Whys
1. **Why** did the service go down? → [proximate cause]
2. **Why** did that happen? → [intermediate cause]
3. **Why** did that happen? → [deeper cause]
4. **Why** did that happen? → [systemic cause]
5. **Why** did that happen? → [ROOT CAUSE]

### Technical Details
[Detailed technical explanation with evidence: logs, metrics, code references]

## Contributing Factors
- [Factor that amplified the impact but isn't the root cause]
- [Factor that delayed detection or response]

## What Went Well
- [Things that worked during the incident]
- [Effective responses or tools]

## What Could Be Improved
- [Detection gaps]
- [Response delays]
- [Communication issues]
- [Tool limitations]

## Action Items

| # | Action | Type | Owner | Priority | Due | Status |
|---|--------|------|-------|----------|-----|--------|
| 1 | [Specific fix for root cause] | Prevent | [name] | P1 | [date] | Open |
| 2 | [Detection improvement] | Detect | [name] | P1 | [date] | Open |
| 3 | [Process improvement] | Process | [name] | P2 | [date] | Open |
| 4 | [Runbook update] | Document | [name] | P2 | [date] | Open |

## Lessons Learned
- [Systemic insight applicable beyond this incident]
- [Pattern to watch for in similar systems]

Postmortem rules:

  • Focus on systems, not individuals
  • Every P1/P2 gets a postmortem within 5 business days
  • Action items have owners, priorities, and deadlines
  • Close postmortem only when all P1 action items are complete
  • Share with entire engineering org to spread learnings

Capacity Planning Template

## Capacity Plan: [Service Name]

**Current baseline** (date: YYYY-MM-DD):
- Peak QPS: [value]
- Average CPU utilization at peak: [%]
- Average memory utilization at peak: [%]
- Current replica count: [N]
- Current node count: [N]

**Growth projection**:
- Monthly growth rate: [%]
- Projected peak QPS in 3 months: [value]
- Projected peak QPS in 6 months: [value]
- Seasonal multiplier: [e.g., 3x for Black Friday]

**Scaling limits**:
- Max replicas (HPA): [N]
- Max nodes (Karpenter/ASG): [N]
- Database max connections: [N]
- Known bottleneck: [component] at [threshold]

**Action required by [date]**:
- [ ] Increase [resource] from [current] to [target]
- [ ] Optimize [bottleneck] before reaching [threshold]
- [ ] Load test at [projected peak * 1.5]

Chaos Engineering Playbook

Pre-Requisites

  • SLOs defined and monitored
  • Error budget gate — the same thresholds as the error_budget_policy above:
    • >50% remaining → run experiments normally
    • 20–50% remaining → only with explicit team approval
    • <20% remaining → do not run chaos experiments; the budget is already burning
  • Runbooks exist for expected failure modes
  • Team has agreed to the experiment scope

Experiment Template

## Chaos Experiment: [Title]

**Hypothesis**: If [failure condition], the system should [expected behavior] within [time].

**Steady state**: [metrics that define "normal" — SLIs within SLO]

**Method**: [What to break]
- Scope: [namespace/service/node]
- Duration: [how long]
- Blast radius: [what's affected]

**Abort conditions**: [when to stop immediately]
- SLI drops below [threshold]
- Error rate exceeds [%]
- Customer reports

**Results**:
- Hypothesis confirmed? [yes/no]
- Unexpected behavior: [description]
- Action items: [list]

Common Experiments

ExperimentToolTests
Kill random podkubectl delete podAuto-recovery, PDB enforcement
Inject latencyChaos Mesh / LitmusCircuit breakers, timeouts, retries
Network partitionChaos MeshGraceful degradation, fallbacks
Fill diskdd if=/dev/zeroAlerts, log rotation, eviction
DNS failureBlock DNS resolutionCaching, error handling, fallbacks
Dependency unavailableBlock egress to serviceCircuit breaker, cached responses

Game Day Planning

  1. Schedule: 2 hours, business hours, all relevant engineers present
  2. Scope: One service or failure domain per game day
  3. Run: Execute experiments, observe responses, take notes
  4. Debrief: Immediately after — what worked, what didn't, action items
  5. Cadence: Monthly for critical services, quarterly for others

Toil Reduction Framework

Identify Toil

Toil is work that is: manual, repetitive, automatable, reactive, scales with service growth, has no enduring value.

Catalog Template

TaskFrequencyDurationAutomatable?Priority
Restart pods after deployDaily10 minYesHigh
Rotate certificatesMonthly30 minYes (cert-manager)High
Clean up old logsWeekly15 minYes (logrotate)Medium
Review alertsDaily20 minPartially (tune alerts)Medium
Manual scalingVariable15 minYes (HPA/Karpenter)High

Target

  • SREs spend <50% time on toil, >50% on engineering projects
  • Track toil hours per sprint
  • Prioritize automation by: frequency * duration * error_risk

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.