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.

Burn Rate Alerts

Burn rate = how fast the error budget is being consumed relative to the budget window. A burn rate of 1 = consuming the budget at exactly the rate that would exhaust it over 28 days.

Alert implementation: [Link to alert config in monitoring tool — e.g. Datadog, Prometheus/Alertmanager, Grafana]

Error Budget Policy

This policy defines what to do with the error budget — both when it's healthy and when it's burning.

When budget is healthy (>50% remaining)

• Feature development and deployments proceed at normal pace
• The team may take on riskier experiments
• Reliability improvements are scheduled but not urgent

When budget is at risk (25–50% remaining)

• Deployment frequency reduced — team ships only well-tested changes
• One reliability improvement added to current sprint
• Weekly error budget review added to team standup

When budget is nearly exhausted (<25% remaining)

• Feature work paused in favour of reliability improvements
• No new deployments without explicit on-call approval
• Daily review of error budget burn rate
• CSM / support notified to manage customer expectations

When budget is exhausted (0% remaining — SLO breached)

• All feature work stops
• On-call engineer and engineering manager notified immediately
• Post-incident review (PIR) required within 5 business days
• SLO target may be temporarily relaxed (with stakeholder approval) while root cause is addressed

Dashboard and Reporting

SLO dashboard: [Link to Datadog / Grafana / etc. dashboard]

Metrics exposed:

• Current SLO compliance (rolling 28-day)
• Error budget remaining (% and minutes)
• Burn rate (current and trend)
• Incident count and MTTR this window

Reporting cadence:

Exclusions and Edge Cases

Planned maintenance: Error budget is not consumed during pre-announced maintenance windows. Maintenance must be communicated [X hours] in advance via [channel].

Dependency failures: If SLO breach is caused by an upstream dependency outside our control, document it — but it still counts against our error budget (our users don't distinguish between our failures and our dependencies' failures).

Force majeure: [Policy for cloud provider outages, major infrastructure events]

SLO Review Cadence

When to change the SLO target:

• Historical baseline has improved significantly and target no longer reflects real reliability
• User feedback indicates the target is misaligned with what users actually experience
• The SLO is being gamed (metric is healthy but users are unhappy)

Quality Checks

• SLIs are user-facing — they measure what users experience, not internal system metrics
• Good and bad events are precisely defined — no ambiguity about what counts
• Targets are based on historical data, not aspirational round numbers
• Error budget policy has clear triggers and clear actions — not "discuss as a team"
• Burn rate alerts have different windows to catch both fast burns and slow burns
• Exclusions are documented so they don't silently inflate the SLO number

Anti-Patterns

• Do not set SLO targets at 100% — this discourages feature development and does not reflect how users experience reliability
• Do not measure internal system metrics as SLIs — SLIs must reflect what users directly experience, not internal CPU or memory
• Do not write an error budget policy with vague triggers — "discuss as a team" is not an actionable policy; triggers must be specific percentages
• Do not base targets on aspirational round numbers — always derive from historical baseline data
• Do not configure only one burn-rate alert window — a single window misses both fast burns and slow burns that exhaust the budget quietly