68 / 95 · 05 Performance & Chaos Engineering · SLOs, SLIs, and Error Budgets← prev⊞ allnext →☰ Read as one page
11.3Choosing Good SLIs
Not all metrics make good SLIs. The best SLIs are:
- User-facing. Measure what users experience, not what servers report.
- Measurable. You must be able to collect the data reliably.
- Actionable. When the SLI degrades, the team can do something about it.
SLI Categories
| Category | Good SLIs | Bad SLIs |
|---|---|---|
| Availability | Successful requests / total requests | Server uptime |
| Latency | p99 request duration | Average response time |
| Quality | Responses with correct content / total responses | Test pass rate |
| Freshness | Data updated within threshold / total queries | Cron job success rate |
| Throughput | Requests served at target rate / total minutes | CPU utilization |
Why "server uptime" is a bad SLI: A server can be "up" (responding to health checks) while returning errors to every user request. Uptime measures infrastructure, not user experience.
Why "average response time" is a bad SLI: Averages hide outliers. If 99% of requests take 100ms and 1% take 30 seconds, the average is ~400ms -- which looks fine but masks a terrible experience for 1% of users.