Most teams can tell you the platform is "up". Far fewer can tell you what "up" means precisely enough to make a decision with, which is the difference between reliability as a plan and reliability as a hope. Three numbers close that gap.
The SLI: what you measure
A service level indicator is a number that describes how well the service is behaving. The good ones are ratios of good events to total events over a window:
- the share of checkout requests that returned successfully in the last 28 days,
- the share of API calls that completed in under 400 milliseconds,
- the share of scheduled jobs that finished before their deadline.
The common mistake is measuring the infrastructure instead of the service. CPU utilisation is not an SLI. A user does not experience CPU; they experience a slow checkout. Measure the thing the user would notice.
The SLO: what you promise
A service level objective is a target for that indicator over a window: 99.9% of checkout requests succeed over 28 days. The number is a business decision, not an engineering one, because it decides how much reliability you are willing to pay for.
99.9% allows about 40 minutes of failure a month. 99.99% allows four. The second one costs substantially more than ten times the first, because it rules out single-region deployments, casual deploys, and most of the ways a small team works. Choosing 99.99% for a product with ten customers is not ambition; it is a decision to spend engineering time on something nobody asked for.
The error budget: what changes behaviour
The error budget is the amount of failure the SLO permits. At 99.9% over 28 days, it is that 0.1% — about 40 minutes. It is the most useful of the three numbers, because it turns arguments about risk into arithmetic.
- Budget remaining. Ship. Take the risk, run the migration, deploy on Friday if you like.
- Budget nearly spent. Slow down. Reliability work takes priority over new features until the budget recovers.
- Budget gone. Freeze. The next change is a fix, not a feature.
That policy is the point. Without it, "we should be more reliable" is an opinion that loses to whatever the roadmap says, every time.
Where teams get it wrong
- Too many SLOs. Three or four per service is plenty. A dashboard of forty is a dashboard nobody reads.
- SLOs nobody agreed to. If the product owner did not sign the number, it is an engineering wish, and it will be overruled the first time it is inconvenient.
- Alerting on everything. Every alert should be tied to a symptom that threatens an SLO. An alert that has never required action is training people to ignore alerts.
- No review. The numbers only change behaviour if someone looks at them on a schedule and says so out loud.
None of this requires new tooling. Prometheus, Grafana, or the monitoring you already pay for can hold all three numbers. What it requires is somebody deciding what "up" means, writing it down, and treating the answer as a commitment.