Availability is advertised as a percentage with a growing number of nines. The figure is precise, but what it measures is narrower than most readers assume.
The number describes a measurement, not an experience
An availability figure is produced by a defined health check running at a defined interval against a defined endpoint. Everything outside that definition is invisible to the calculation.
A service that returns errors for a subset of requests, or responds correctly but far too slowly, can register as fully available throughout. The check passed, so the minute counted.
This is why customer reports and provider dashboards diverge during incidents. Both are accurate about different things.
Nines are budgets of time
Three nines permits roughly forty minutes of downtime per month, four nines around four minutes, and five nines a few seconds. Each additional nine reduces the allowance tenfold.
Because the budget shrinks so fast, the difference between four and five nines is not an incremental engineering effort. It usually requires removing manual intervention from the recovery path entirely.
A human being paged, waking up and connecting already exceeds a five-nines monthly budget, which is why that tier implies automated failover rather than better operations.
Composite systems multiply
An application depending on several services inherits the availability of all of them. Where the dependencies are sequential, the individual figures multiply and the result is lower than any single component.
Ten dependencies at three nines each produce an application substantially worse than three nines, without anything having failed unexpectedly.
Redundancy reverses the arithmetic, but only where the redundant paths fail independently. Shared power, shared network or shared configuration reintroduce correlation and undo the benefit.
The credit is not compensation
Service level agreements typically remedy breaches with a credit against future spending, proportional to the shortfall rather than to the harm caused.
Claiming usually requires the customer to submit evidence within a window, so credits are not applied automatically. Many go unclaimed.
Read the exclusions before the number
Scheduled maintenance is generally excluded, as are failures attributed to customer configuration, third-party networks or events outside the provider's control.
Those exclusions remove a meaningful share of real-world unavailability from the calculation, which is part of why measured figures usually exceed what users perceive.
The practical reading is that the percentage describes the provider's obligations rather than predicting how often an application will work, and only the surrounding definitions reveal the difference.