An availability target of 99.9 per cent permits eight hours and forty-six minutes of downtime a year. That is a full working day, and it is the level most commercial agreements offer — which makes the phrase "three nines" considerably less reassuring than it sounds.
Each additional nine divides the allowance by ten, and the cost of achieving it does not divide by anything.
What does each level allow?
The arithmetic is simple and the numbers are worth memorising.
| Availability | Per year | Per month |
|---|---|---|
| 99% | 3 days 15 h | 7 h 12 m |
| 99.9% | 8 h 46 m | 43 m 12 s |
| 99.95% | 4 h 23 m | 21 m 36 s |
| 99.99% | 52 m 36 s | 4 m 19 s |
| 99.999% | 5 m 16 s | 26 s |
The monthly column is the one that matters in practice, because most agreements are measured and credited monthly. A single incident lasting an hour breaches a monthly three-nines target outright, with no way to earn it back within the window.
The annual figure allows the same incident to be averaged against eleven good months, which is why the measurement window is not a detail. The same target is a different promise depending on it.
Why does nobody offer 100 per cent?
Because it cannot be delivered and would be a commitment to compensate for every failure including ones outside the provider’s control. Network paths, upstream providers and physical infrastructure all fail occasionally.
The honest framing is that availability is bought in increments and each one costs more than the last. Going from two nines to three is usually a matter of competent operations; going from four to five means redundancy at every layer, and the last nine can cost more than everything before it.
That is why the right question is not how many nines are possible but how much downtime the thing being protected can actually tolerate.
What is an error budget?
The allowance restated as something to spend. If the target is 99.9 per cent, the budget for the month is 43 minutes, and every minute of downtime consumes part of it.
The value of framing it that way is that it stops availability being a binary. A team with budget remaining can take risks — ship faster, run a migration, try a change — and a team that has spent it should be stabilising instead.
It also makes the trade explicit. Unused budget is not a prize; it can mean the target is set too conservatively for what the service actually needs.
What counts as down?
Whatever the agreement says, and the definition does more work than the percentage. A service can be responding to health checks while being useless to every real user, and if the measurement is the health check it is recorded as available.
Partial failures are the hard case. A service working for most users and failing for one region, or working for reads and failing for writes, is not fully down and not fully up, and how it is counted has to be written down in advance.
Who measures matters as much. A provider measuring from inside its own network reports better numbers than a customer measuring from outside it, and both are measuring honestly.
What do service credits actually cover?
A portion of the fee, and nothing else. A breach typically entitles a customer to a percentage of the monthly charge back — which is a refund for service not delivered rather than compensation for what the outage cost the business.
That asymmetry is deliberate and normal. No provider can underwrite its customers’ losses, so the credit is a signal about seriousness rather than an insurance payout.
The practical implication is that an availability target is not risk transfer. Anything whose failure would be expensive needs its own resilience, regardless of what the agreement promises.
How is availability measured over a period?
As uptime divided by total time in the window, which means the window has to be stated. The same incident produces a different percentage against a month, a quarter or a year, and quoting a figure without its period says almost nothing.
Rolling windows behave differently again. A rolling thirty days moves the breach point continuously rather than resetting on the first of the month, which is stricter and harder to game.
Questions people ask
Does a status page count as evidence? It is the provider’s own record. An independent measurement is what settles a disagreement.
Can I claim credits automatically? Rarely. Most agreements require the customer to notice and to ask within a window.
Is 99.5 per cent a reasonable target? For an internal tool, often. It permits about three and a half hours a month, which is generous enough to be achievable without redundancy.
Is planned maintenance counted? Usually excluded, which is worth checking. An agreement excluding it is promising less than it appears to.
Does availability mean reliability? No. A service can be available and slow, or available and returning wrong answers.
How do I combine two dependencies? For services in series, multiply their availabilities. Two at 99.9 per cent give 99.8 together.
What target should I choose? The one matching the cost of downtime. Choosing more nines than the business needs is a way to spend money on nothing.
Do the arithmetic before agreeing to the number. The uptime percentage calculator converts between a target and the downtime it permits over any window, and works the other way from an incident log to the availability it produced.