Availability Nines

In practice

Availability	Per 30 days	Per year
99%	7h 12m	3d 15h 36m
99.9%	43m 12s	8h 45m 36s
99.95%	21m 36s	4h 22m 48s
99.99%	4m 19.2s	52m 34s
99.999%	25.9s	5m 15s

Two facts that reshape the conversation:

Serial dependencies multiply. A request passing through five components that are each 99.9% available is 0.999^5 = 99.5% — three nines of components gives you two-and-a-half nines of system. Adding a component costs availability unless it is optional.

Redundancy adds nines only if failures are independent. Two 99.9% replicas give 99.9999% if they never fail together — and they share a region, a deploy pipeline, a config store and a certificate expiry. Correlated failure is why real multi-AZ systems land nearer 99.99% than the arithmetic promises.

When it matters

Setting an SLO, sizing redundancy, and pricing an SLA before it is signed.

Common mistake

Quoting nines without a window. 99.9% per year tolerates a single 8-hour outage; 99.9% per 30 days does not. The window is half the commitment.