SLI, SLO and error budget

In short: An SLI measures the quality of a service (for example share of successful requests), an SLO is the target for it (for example 99.9 %), the error budget the permitted amount of errors that is left until then.

In more detail: Instead of “100 % availability” (unaffordable) you set a sensible target. At 99.9 % over 30 days about 43 minutes of downtime are allowed. When the budget is used up, stability takes priority over new features.

In Depth

• SLI: Availability, latency (for example 95th percentile), error rate, throughput.
• SLO: Internally set target. SLA: Contract with customers, often with penalties.
• Error budget: 100 % − SLO. Steers risk: budget left → ship faster; empty → freeze.
• Percentiles instead of average: Averages hide slow outliers.
• Burn-rate alerts: Alert when the budget is used up too fast.
• Post-mortem: Blameless review after incidents.

See also: Observability: logs, metrics and traces, Circuit breaker and backpressure