SLOs and Error Budgets in Practice

This is for you if

• You've been asked to set up SLOs and want more than a dashboard
• You run Prometheus and want alerts that page only when users hurt
• You're preparing for SRE or platform interviews
• You own a data pipeline or LLM feature and need real reliability targets

What's inside

• Choosing SLIs from critical user journeys, with specification vs implementation
• The nines table and error budget math with worked examples, all checked
• Multi-window multi-burn-rate alerts derived from first principles (14.4, 6, 3, 1)
• Detection and reset time math, plus low-traffic strategies
• Layered Prometheus 3 recording rules, including the missing-numerator fix
• promtool unit tests that prove when alerts fire and clear
• What Sloth and Pyrra generate, line by line, and how they differ
• SLOs for batch pipelines (freshness, completeness) and LLM features (valid output, TTFT, quality)
• Error budget policy, SLO review template, troubleshooting playbook and 20 interview Q&As

After this guide you can

• Turn a user journey into SLIs and a defensible SLO with a written rationale
• Compute budgets, burn rates, detection times and incident costs without guessing
• Implement and unit test burn-rate alerts in Prometheus, by hand or with Sloth or Pyrra
• Define freshness and LLM SLOs that catch failures HTTP codes miss
• Write an error budget policy and run an SLO review that changes decisions

Contents

1. The mental modelp. 3
2. Choosing SLIs from user journeysp. 10
3. Writing SLOs and doing the error budget mathp. 15
4. Implementing SLIs with Prometheus recording rulesp. 22
5. Multi-window, multi-burn-rate alertingp. 29
6. Let the tools write the rules: Sloth and Pyrrap. 39
7. SLOs for batch pipelines and AI featuresp. 47
8. Error budget policies, SLO reviews and buy-inp. 55
9. Troubleshooting playbook and common mistakesp. 59
10. Production checklist and design decisionsp. 62
11. Interview questionsp. 64
12. Cheat sheetp. 70
13. Your 30-day practice planp. 73