---
description: SLOs, SLIs, and error budgets—definitions, choosing SLIs, error budget policy, burn-rate alerting. Balance reliability and velocity.
alwaysApply: false
---

# SLOs, SLIs, and Error Budgets

Guidelines for service level objectives and error budgets.

## Definitions

- **SLI**: Quantitative measure of service behavior (e.g. successful_requests / total_requests, or requests_under_500ms / total). Must be measurable and user-relevant.
- **SLO**: Target for an SLI (e.g. 99.9% availability over 30 days). Internal target; often stricter than SLA.
- **SLA**: Contract with consequences (e.g. credits); usually looser than SLO to allow buffer.
- **Error budget**: 1 − SLO (e.g. 0.1% = ~43 min/month). Consumed by outages and bad releases; when exhausted, prioritize reliability over features.

## Choosing SLIs

- **Availability**: Success / total (HTTP 2xx/3xx, or health checks). Good for APIs and user-facing services.
- **Latency**: Requests under threshold / total (e.g. p99 < 500ms). Good for responsiveness.
- **Throughput**: Request rate or success rate; good for batch or async.
- **Quality**: Optional (e.g. % correct results). Use sparingly; keep SLI count small (2–4 per service).

## Error Budget Policy

- **Healthy (>50% budget)**: Normal velocity; experiments allowed.
- **Caution (25–50%)**: Review changes; limit risky deploys.
- **Critical (10–25%)**: Focus on reliability; rollback plans required.
- **Exhausted (<10%)**: Feature freeze; all-hands reliability; postmortem before deploy.

## Alerting on SLOs

- Use **burn rate** (how fast budget is consumed) rather than raw error rate. Multi-window (e.g. 1h and 6h) reduces false positives. Alert when burn rate would exhaust budget within the window.
- Link alerts to runbooks; severity by impact on budget.

## Definition of Done (SLO)

- [ ] SLI defined and measured; SLO target and window set.
- [ ] Error budget policy documented and communicated.
- [ ] Alerts and dashboards in place; runbooks linked.

## Common Pitfalls

- **SLO as exact target** - SLO defines acceptable reliability; use error budget for tradeoffs, not “hit exactly 99.9%.”
- **Too many SLOs** - Focus on 2–4 key SLIs; avoid vanity metrics.
- **No consequences** - Error budget must influence prioritization (freeze, rollback, focus on reliability).