Availability targets and budgets
The arithmetic on this page is the arithmetic in Nines, the calculator built for exactly this. Every number shows the formula that produced it.
The target
99.9% availability, measured monthly, per host.
Three nines. It is chosen to be honest for a static site on a managed host with no on-call rotation, not to be impressive. A target nobody can hold is worse than a modest one that means something.
What 99.9% allows
The formula:
allowed = window x (1 - target)
| Window | Length | Allowed downtime |
|---|---|---|
| per day | 1440 min | 1.44 minutes |
| per week | 10080 min | 10.08 minutes |
| a 30 day month | 43200 min | 43.2 minutes |
| a 31 day month | 44640 min | 44.64 minutes |
| a 365 day year | 525600 min | 8.76 hours |
Month lengths are real. A 30 day month and a 31 day month differ by 1.44 minutes of allowed downtime at this target, and rounding every month to 30 days quietly gives away budget in the long ones and overspends it in the short.
The error budget
The budget is the allowed downtime. Spending it is logging an outage against it.
burn rate = (spent / days elapsed) / (budget / days in window)
A burn rate of 1.00x is exactly on pace: the budget runs out precisely as the month ends. Above 1.00x it runs out early, and the date it runs out is stated rather than left as an exercise.
Current budget
No outages have been logged against this target as of 2026-09-21.
That is not a claim of perfect availability. It is a claim that no outage has been recorded, which is a weaker and truer statement. There is no monitor feeding this page; see the status page for what is and is not measured.
When an outage is recorded it will appear here with its start, its end, and the minutes it cost.
Composite availability
Most things depend on other things, and dependencies in series multiply:
A = A1 x A2 x ... x An
Two dependencies at 99.9% in series give 99.8001%, which is worse than either. That is the part people are surprised by: a chain is always worse than its weakest link, never equal to it.
Redundancy works the other way, multiplying the unavailabilities:
A = 1 - (1 - A1) x (1 - A2) x ... x (1 - An)
Two independent 99% paths in parallel give 99.99%, better than either alone, as long as they really are independent.
Latency
Same arithmetic, requests in place of minutes. A p99 target does not promise every request is fast; it allows one in a hundred to miss the threshold. There is no latency SLO recorded for this network today, because there is no request-level measurement to hold one against.