Loading Uptime & SLA Calculator…
Anything from 90 to 99.9999.
28, 30 or 31 — an SLA month is a billing month, not an average.
43m 12s
2,592 seconds out of 2,592,000
The error budget is the same sum read backwards.
Adding one more nine takes the target to 99.99% and divides the allowance by ten: 43m 12s becomes 4m 19s per month. That is the real cost of a nine, and it is paid in redundancy and on-call rather than in effort.
Two published tables can disagree by a couple of seconds and both be right: a year of 365 days gives four nines an allowance of 52m 34s, while the 365.25-day year some calculators use gives 52m 36s. The gap is smaller than the monitoring interval that would detect it, which is a useful reminder that these numbers are contract arithmetic, not measurement.
Uptime targets are quoted in nines because the arithmetic is uncomfortable in minutes. Over a 365-day year, 99% permits 3d 15h 36m of downtime. Three nines cuts that to 8h 45m 36s, four nines to 52m 34s, and five nines to 5m 15s. Each additional nine divides the allowance by exactly ten, which is why the step from three to four is the one that changes how a team is staffed rather than how carefully it deploys.
The in-between figures are worth knowing too. A 99.5% target — common in small print — still allows 1d 19h 48m a year, and 99.95% allows 4h 22m 48s. Anything below three nines is a promise that a working day can disappear.
A percentage means nothing without the window it is measured over, and the same 99.9% target reads very differently at each scale: 1m 26s in a day, 10m 5s in a week, 43m 12s in a 30-day month, 2h 11m 24s in a quarter.
Most contracts measure a billing period rather than a rolling window, and that is the detail that catches people out. One four-hour incident on a Tuesday afternoon puts a 30-day month at 99.44%, which misses three nines by a distance no amount of good behaviour later in the month can recover — the clock only resets when the next invoice starts. A rolling 30-day window would have let the same outage age out gradually; a calendar month does not.
The length of the month matters as well, which is why it is an input here rather than an average. At 99.9%, February allows 40m 19s while a 31-day month allows 44m 38s — a spread of more than four minutes on the same target.
If a 99.9% target permits 43m 12s in a 30-day month, then 43m 12s is a budget you are allowed to spend. Twelve minutes of downtime leaves 31m 12s, or 28% of the month consumed. Nothing in the arithmetic changes; only the direction of the subtraction does.
What changes is the conversation. An allowance is something to apologise for using; a budget is something to spend deliberately, on a risky migration or a load test in production. Teams watching the remaining figure ship faster while it is healthy and freeze when it is nearly gone, which beats caution applied evenly all month.
These figures are contract arithmetic, not measurement. A 40-second outage falling between two 60-second health checks may never appear in the uptime figure, and a service answering in eight seconds is still counted as available though nobody would call it working. Most agreements also exclude announced maintenance and anything upstream, so the measured figure is kinder than the experience.
Two published tables can disagree by a couple of seconds and both be right. Using a 365-day year, four nines allows 52m 34s; using the 365.25-day year some calculators prefer, it allows 52m 36s. The gap is smaller than most monitoring intervals — a reminder that the last digit of an availability figure is a rounding convention rather than a fact about the service.
Just under 22 minutes: 21m 36s across a 30-day month, or 22m 19s in a 31-day one. That is the target most managed platforms offer on a standard plan, and it means a single failed deploy that takes half an hour to roll back has already broken it.
Usually not, provided the window was announced in advance and inside the notice period the contract specifies. That exclusion is worth reading closely, because a provider with generous maintenance terms can report a far better figure than a customer experienced. Emergency maintenance is treated as an outage in most agreements.
Not when the SLA measures a billing period, which is the usual case. Each month is assessed on its own, so a single bad afternoon can cost the target for that month while the annual average still looks excellent. Averaging across months is only valid where the contract explicitly says the term is a year.
Typically a service credit against the next invoice on a sliding scale — often 10% of the monthly fee for a small miss and 25% or more for a large one. Credits are calculated on what you paid rather than on what the outage cost you, and most providers require a claim within a set number of days.
Both conventions exist, and they produce different numbers from identical behaviour. Time-based measurement counts intervals in which probes failed; request-based divides successful requests by total requests. A brief outage in your quietest hour looks worse under the first and better under the second.