Slo Error Budget

SLO and Error Budget Skill

Produce a complete, implementable SLO document for a service — covering what to measure, what target to set, how to calculate the error budget, and what to do when it burns.

A good SLO is not a target to hit. It is an agreement about what reliability means for your users — and a framework for making principled trade-offs between reliability and velocity.

Where this sits — the frame of the spine

This is the governor of the incident-response spine: slo-error-budget (frame) → /debugging-log-analyser → /incident-postmortem → /oncall-runbook. It sets the error budget the whole loop runs inside — the objective forcing function that later decides whether a postmortem's action items get done now (budget spent) or deferred (budget healthy). Shared terms (SLO, error budget, incident, action item) are defined once in docs/craft/incident-response.md.

The loop

An SLO fails when it's aspirational instead of user-derived, or when the budget has no teeth. Phase 1 is load-bearing: a target picked from "100% minus a bit" defends nothing.

1. Derive the target from users, not from 100%. Pick the SLIs that track what users actually feel (success rate, latency, freshness) and a target from what they need — the point where more reliability stops mattering to them. Done when: each SLO target traces to a stated user need, and none is 100% or a round number chosen for looking good.
2. Turn the target into a budget with math. Compute the error budget (100% minus the SLO, over the window) as a concrete allowance — requests, minutes, or events — not a percentage nobody feels. Done when: the budget is expressed as a countable allowance for the window, and burn-rate alerts fire before it's spent, not after.
3. Give the budget teeth — the policy is the point. Write what changes at each budget level: healthy → ship; at risk → slow and shore up; exhausted → stop feature work and fix reliability. A budget with no policy is a dashboard nobody obeys. Done when: each budget level names a specific, enforced consequence — including a real feature-freeze trigger — that a team would actually follow.
4. Hand off as the loop's governor. State that this budget is what prioritises /incident-postmortem action items: budget spent makes them urgent, budget healthy lets them wait. Done when: the loop downstream could use this budget to decide action-item urgency without re-litigating the reliability target.

Required Inputs

Ask for these if not already provided:

• Service name and brief description of what it does
• Primary users — who depends on this service and how
• User-facing interactions to protect — e.g. API calls, page loads, transactions
• Current reliability data — error rate, latency, uptime (last 30–90 days if available)
• Existing on-call setup — who responds to alerts?
• Deployment frequency — how often does the team ship?
• Any existing SLAs with customers — these constrain SLO targets

Key Definitions

Always establish these before writing the SLO:

Output Format

SLO Document: [Service Name]

Service: [Name] | Team: [Team name] Owner: [Name / role] | Approved by: [Name] Effective date: [Date] | Review date: [Date + 3 months] Version: [1.0]

Why This SLO Exists

[2–3 sentences. What reliability problem are we solving? What was happening before this SLO that made us need it? What decision-making does this SLO enable?]

Service Overview

What this service does: [One sentence] Who depends on it: [Internal teams / external customers / both — describe] Critical user journeys protected by this SLO:

1. [Journey 1 — e.g. "User completes a payment"]
2. [Journey 2]
3. [Journey 3]

SLIs — What We Measure

Define one SLI per user journey or reliability dimension. Keep it to 3–5 SLIs maximum.

SLI 1: [Name — e.g. Request Success Rate]

SLI 2: [Name — e.g. Latency]

SLI 3: [Name — e.g. Data Freshness / Queue Depth / etc.]

[Same structure]

SLO Targets

How targets were set:

• Historical baseline (last 90 days): [X]%
• Target is set [above / at] historical baseline to [improve reliability / reflect current reality while formalising the commitment]
• Rationale: [1–2 sentences]

What 100% is NOT the target: [Brief explanation of why targeting 100% is counterproductive — it discourages feature development and doesn't reflect user reality]

Error Budget Calculation

For SLI 1 ([Name]), at [X]% target:

Error budget = (100% - SLO target) × measurement window
             = (100% - [X]%) × 28 days × 24 hours × 60 minutes
             = [Y]% × [Z total minutes]
             = [N] minutes of allowed failure per 28-day window

In plain terms: We can afford [N] minutes of [bad events] in any rolling 28-day window before we breach the SLO.