An SLO allows a bounded amount of unsuccessful service. That allowance is the error budget.

If an availability SLO is $99.9%$, the corresponding unavailable fraction is

$$1-0.999=0.001,$$

or $0.1%$ of the measured opportunity over the SLO window.

The budget can be consumed by failed requests, excessive latency or another event defined by the SLI.

This gives reliability work a quantitative connection to product change. When the service is comfortably within its objective, the team can accept some deployment risk. When the budget is being exhausted rapidly, slowing risky changes and prioritizing reliability may be rational.

The budget is not permission to intentionally cause failures. It acknowledges that pursuing perfect reliability has increasing cost and can prevent useful change.

Error budgets turn an SLO from a passive target into a shared way to reason about the trade-off between reliability and delivery velocity.