Convert an availability target into allowed downtime.
Allowed downtime
Per year
8h 45m
Per month
43m 50s
Per week
10m 5s
Per day
1m 26s
This is the maximum downtime budget for a single service at that availability. When services depend on each other in series, their availabilities multiply, so a chain is always less available than its weakest link — a system of five 99.9% services is only ~99.5% end to end.
Enter an availability target — or pick a preset (two to five nines) — to see the maximum downtime you are allowing per year, month, week, and day. The gap looks small on paper but is huge in practice: 99.9% permits about 8 hours 45 minutes of downtime a year, while 99.99% allows only about 52 minutes. That order-of-magnitude difference is exactly why SLAs are stated in nines.
Use it to sanity-check an SLA before you commit to it, and to reason about the engineering it implies — three nines is achievable with good deploys and monitoring, four or five nines demands redundancy, automated failover, and error budgets. Remember that dependent services multiply: a chain of components is always less available than its weakest link.
About 8 hours 45 minutes per year, or roughly 43 minutes per month. Each additional nine cuts that by ~10×: 99.99% allows about 52 minutes per year, and 99.999% ("five nines") allows only about 5 minutes per year.
For services in series (each must be up for the request to succeed), multiply their availabilities. Five services at 99.9% each give 0.999^5 ≈ 99.5% overall — noticeably worse than any single component. Redundancy (parallel paths) works the opposite way and raises combined availability.