Avoiding Goodhart's Law - Use SLO's as Tools Not Cudgels
summary
Argues that SLIs/SLOs must measure outcomes not be punitive targets, and shows concrete fixes: an SLA that refunds when queue length >100 teaches gaming so the team replaces that with a nested set of SLIs, an APM-backed single source of SLI truth (Datadog / New Relic / Honeycomb / Instana), percentile-based metrics, a 28-day sliding window SLO, a repeat-counter for client retries (beyond 10s) and an 8-hour SLA reaction — all negotiated against code, infrastructure and customer experience.
AI-written summary · about
chapters (9)
- 0:00SLOs and Goodhart's Law
- 3:45Goodhart's Law experience
- 8:01Avoiding Goodhart's Law with SLIs
- 10:50Three dimensions for SLIs and SLOs
- 16:29Code SLIs/SLOs
- 17:10Infrastructure SLIs/SLOs
- 21:21Business and customer experience
- 25:13Negotiating SLAs
- 32:15Key takeaways
related talks
06ma pint sized introduction to SLO
30mMetrics for good, not evil
31m9 Things You Should Do When Starting to Use SLOs
01h 06mPepperoni, Puppets, and Priorities: Demystifying SLAs, SLOs, and SLIs
38mImplementing SRE practices: SLI/SLO deep dive
33mHow and why we lowered our SLO
09mHow to answer tricky questions the SRE way
30mTechniques for SLOs and Error Budgets at Scale