Write an on-call runbook for a service

On-Call Runbook Skill

Produce a complete on-call runbook for a service — giving the on-call engineer everything they need to respond confidently to alerts at 3am, without having to ask anyone for help.

A good on-call runbook reduces mean time to resolution (MTTR) by eliminating the "what do I do first?" problem. It is written for the on-call engineer who has just been paged and needs to act, not for someone calmly reading documentation.

Where this sits — the spine's terminus

Last in the incident-response spine: /slo-error-budget (frame) → /debugging-log-analyser → /incident-postmortem → oncall-runbook. It receives the contributing factors and action items from /incident-postmortem and turns the detection/mitigation learnings into an entry that makes the next responder minutes, not hours — closing the loop so the same incident doesn't recur at full cost. Runbook entry, detection/mitigation time, and the loop are defined once in docs/craft/incident-response.md.

The loop

A runbook fails when it's written for a calm reader instead of a paged one at 3am. Phase 1 sets the audience; every later choice serves it.

1. Write for the paged engineer, not the documentarian. The reader has just been woken and needs to act — so lead with the fastest safe mitigation, put copy-pasteable commands first, and defer background. Prose that explains before it acts fails at 3am. Done when: each alert's entry lets a non-expert take the first safe action within a minute of opening it, without reading theory.
2. Turn postmortem learnings into per-alert procedures. For each known failure (the incident-postmortem's are the highest-value), write detect → mitigate → escalate: the exact checks, the copy-pasteable commands, the rollback, and when to page whom. Done when: every alert maps to a procedure with concrete commands and a clear mitigation, not just "investigate."
3. Make escalation and handoff unambiguous. Who to page, when, and how to hand off mid-incident — because the second failure mode after "what do I do?" is "who do I wake, and when is it okay to?" Done when: the escalation matrix names people/rotations and the trigger for each, and the handoff template captures state so the next responder isn't starting cold.
4. Close the loop back to prevention. Flag where a runbook step reveals a gap that should become monitoring or an /slo-error-budget action — the runbook is where the loop's learnings surface the next prevention. Done when: gaps found while writing the runbook are logged as detection/prevention improvements, not silently absorbed.

Required Inputs

Ask for these if not already provided:

• Service name and what it does
• Team and tech lead name
• Alert list — names of alerts that currently page on-call
• Monitoring setup — Datadog / Grafana / CloudWatch / PagerDuty / etc.
• Common failure modes — what breaks most often, and what fixes it
• Escalation contacts — who to call when on-call can't resolve it
• Deployment setup — can on-call roll back? How?
• Service dependencies — what does this service depend on, and what depends on it?

Output Format

---

On-Call Runbook: [Service Name]

Team: [Team name] | Tech lead: [Name] PagerDuty service: [Link] | Escalation policy: [Policy name] Last updated: [Date] | Next review: [Date + 90 days]

First time on-call for this service? Read the [developer onboarding doc] first — it covers the architecture and how things work. This runbook assumes you understand the service.

---

Quick Reference

Dashboard: [Link — the first thing to open when paged] Logs: [Link — where to find logs] Runbook index: Jump to the alert that paged you → [Alert list below] Can't resolve in 30 min? Escalate to: [Name] via [Slack / PagerDuty]

Rollback command (memorise this):

[rollback command — e.g. kubectl rollout undo deployment/[service-name]]

---

Escalation Matrix

Situation	Escalate to	How	After how long
Can't diagnose the alert	[Tech lead name]	Slack DM / Phone	30 minutes
Alert requires infra change	[Platform team]	#platform Slack	Immediately
Customer-facing impact	[CSM / Support lead]	#incidents Slack	Immediately (P1)
Database issue	[DBA or data team]	Slack / PagerDuty	Immediately
[Specific dependency] down	[[Dependency] on-call]	PagerDuty / Slack	Immediately
Extended outage (>1 hour)	[Engineering manager]	Phone	1 hour

Contacts:

Name	Role	Slack	Phone
[Name]	Tech lead	@[handle]	[Number]
[Name]	Engineering manager	@[handle]	[Number]
[Name]	Platform / infra	@[handle]	[Number]
[Platform team]	Infra on-call	#platform	PagerDuty

---

Service Architecture (Quick View)

[Upstream callers]
        │
        ▼
[This Service]
        │
        ├──→ [Primary Database]
        ├──→ [Cache — e.g. Redis]
        └──→ [Downstream Service / Queue]

If this service is down, these are affected: [List downstream consumers] If these are down, this service is affected: [List upstream dependencies]

---

Alert Runbooks

ALERT: [Alert Name 1 — e.g. HighErrorRate]

What it means: [Plain English — e.g. "More than 5% of API requests are returning 5xx errors in the last 5 minutes"] Severity: P1 / P2 / P3 SLO impact: Yes / No — [If yes: this alert means the error budget is burning at [X]× rate]

Step 1 — Acknowledge and assess

# Check current error rate
[query or dashboard link]

# Check which endpoints are erroring
[query or command]

Step 2 — Check recent changes

# Any deploys in the last hour?
[command or link to deployment log]

# Recent config changes?
[where to check]

Step 3 — Check dependencies

# Is the database healthy?
[health check command or link]

# Is [downstream service] healthy?
[health check command or link]

Step 4 — Diagnose

If you see	It means	Do this
[Error pattern 1]	[Cause]	[Action]
[Error pattern 2]	[Cause]	[Action]
[Error pattern 3]	[Cause]	[Action]
No clear pattern	Unknown cause	Escalate to [name]

Step 5 — Fix or mitigate

# If caused by bad deploy — roll back:
[rollback command]

# If caused by [specific issue]:
[fix command]

# If caused by upstream dependency:
[mitigation — e.g. enable circuit breaker, reduce traffic, etc.]

After resolving:

• Confirm error rate has returned to baseline
• Check no downstream services were affected
• If P1: open a post-incident review — see [incident-postmortem skill]
• Update #incidents with resolution summary

---

ALERT: [Alert Name 2 — e.g. HighLatency]

What it means: [e.g. "P99 response time has exceeded 1s for more than 3 consecutive minutes"] Severity: P1 / P2 / P3 SLO impact: Yes — latency SLO breach

Step 1 — Assess scope

# Check which endpoints are slow
[query or dashboard — broken down by endpoint]

# Check if latency is across all regions or localised
[query or command]

Step 2 — Common causes and fixes

Cause	Signal	Fix
Database slow queries	DB latency spike on dashboard	[Check slow query log: command]
Cache miss storm	Cache hit rate drops on dashboard	[command or action]
Memory pressure / GC	High memory on service dashboard	[command or action — e.g. restart, scale up]
Upstream service slow	Trace shows time in external call	Escalate to [service] on-call
Traffic spike	Request rate spike on dashboard	[Scale up: command]

Step 3 — Escalate if unresolved in 20 minutes Page [Tech lead] via PagerDuty / Slack.

---

ALERT: [Alert Name 3 — e.g. DatabaseConnectionPoolExhausted]

What it means: [e.g. "The service has used all available database connections — new requests will fail"] Severity: P1 SLO impact: Yes — will cause errors immediately

Immediate mitigation:

# Restart the service to flush stale connections
[restart command]

# Check current connection count
[DB connection query]

Diagnose root cause after stabilising:

# Check for long-running queries holding connections
[query]

# Check if a recent deploy changed connection pool config
[where to check]

Resolution: [e.g. "Increase pool size in config / kill long-running queries / scale the service"]

---

ALERT: [Alert Name 4 — e.g. QueueBacklogHigh / ConsumerLag]

What it means: [e.g. "The message queue backlog exceeds 10,000 messages — consumers are not keeping up"] Severity: P2 SLO impact: Depends — if queue backs up, downstream systems will receive delayed data

Step 1 — Check consumer health

# Are consumers running?
[command]

# Consumer error rate?
[dashboard or query]

Step 2 — Check message contents

# Are there poison messages causing retries?
[command to inspect dead-letter queue or failed messages]

Step 3 — Options

If	Then
Consumers are down	Restart consumers: [command]
Poison message in queue	Move to DLQ: [command]
Consumers healthy but slow	Scale consumers: [command]
Upstream producing too fast	Escalate to [upstream service] owner

---

ALERT: [Add additional alerts following the same pattern]

---

Diagnostic Cheat Sheet

Common commands for quick diagnosis. Paste and run without modification.

# Service health
[health check command]

# Recent logs (last 100 lines)
[log command]

# Error logs only
[error log filter command]

# Current pod / instance status
[kubectl get pods / aws ecs describe-tasks / etc.]

# Restart the service
[restart command]

# Roll back to previous version
[rollback command]

# Database connection count
[DB query]

# Cache hit rate
[cache stats command]

# Current request rate
[metrics query]

---

Useful Dashboard Links

Dashboard	URL	Use it to
Service overview	[Link]	First stop — error rate, latency, request rate
Database	[Link]	Connection count, slow queries, replication lag
Infrastructure	[Link]	CPU, memory, disk
Queue / consumers	[Link]	Backlog depth, consumer throughput
Upstream dependencies	[Link]	Dependency health at a glance

---

Incident Communication

When you declare an incident:

Post to #incidents immediately:

🔴 INCIDENT — [Service Name]
Status: Investigating
Impact: [Who is affected and how]
Paged: [Your name]
Next update: [Time — max 30 min from now]

Update every 30 minutes while active:

🔴 UPDATE — [Service Name] — [Time]
Status: [Investigating / Identified / Mitigating / Resolved]
Latest: [One sentence on what you found or did]
Next update: [Time]

On resolution:

✅ RESOLVED — [Service Name] — [Time]
Duration: [X minutes]
Impact: [Summary of who was affected]
Cause: [One sentence]
Follow-up: [PIR required? Yes/No — link when created]

---

On-Call Handoff

Use this template at the end of every on-call shift:

--- ON-CALL HANDOFF: [Service Name] ---
Date: [Date]
Outgoing: [Your name]
Incoming: [Next on-call name]

INCIDENTS THIS SHIFT:
- [Incident summary — date, duration, cause, resolution, follow-up required]

OPEN ISSUES TO WATCH:
- [Anything not fully resolved / trending in the wrong direction]

CHANGES SINCE LAST HANDOFF:
- [Deploys, config changes, infra changes that affect on-call awareness]

RUNBOOK GAPS FOUND:
- [Anything you had to figure out that isn't documented — please add it]

ANYTHING ELSE:
- [Notes for incoming on-call]

---

Quality Checks

• Every alert that pages on-call has a runbook entry — no alert is missing
• Rollback command is accurate and tested recently
• Escalation contacts have current phone numbers and Slack handles
• Diagnostic commands work — they have been run by at least one person recently
• Handoff template is used at every shift change — not just during incidents
• "Things I had to figure out that weren't documented" are added to this runbook after every incident

Anti-Patterns

• Do not write alert runbooks with vague diagnostic steps like "check the logs" — every step must specify the exact command, dashboard link, or query to run
• Do not include an alert in the runbook that has no specific on-call action — an alert that pages someone with no defined response path creates panic, not resolution
• Do not leave the rollback command undocumented or untested — a rollback procedure that has never been run will fail when needed most
• Do not list escalation contacts without phone numbers and Slack handles — email-only escalation paths are useless during a 3am incident
• Do not write the runbook once and treat it as permanent — runbooks go stale after incidents; every incident must trigger a review of the relevant runbook entries