Availability and error budget calculator
Turn a number somebody promised in a meeting into the downtime it actually permits, then find out what happens to it once five services have to work at once.
Runs in your browser. Nothing is sent anywhere.
99.9%
three nines
1.44 min
Per day
10.1 min
Per week
43.8 min
Per month
8.76 hr
Per year
In series
Everything must work
99.501%
5 services at 99.9% each. Real availability is 1.82 days of downtime a year, not 8.76 hr.
In parallel
Any one is enough
99.9999%
Redundancy multiplies the nines, which is why it is the first lever anyone reaches for. It only holds if the copies fail independently.
Error budget
What is left this month
27%
31.8 min of budget left. Spend it deliberately on risky deploys rather than discovering it is gone.
The part worth repeating
5 services at 99.9% each is 99.50%, not 99.9%.
Availability is quoted per component and experienced per request. Every service in the path has to work, so the nines multiply: that is 1.82 days of downtime a year rather than the 8.76 hr anybody was promised. It is the number that changes how the meeting goes.
The card shows the numbers only. Nothing you typed is included.
The number that changes how people argue is the series one. Five services at three nines each, all of which have to work for a request to succeed, gives you about 99.5%, which is roughly forty-four hours of downtime a year rather than nine. Availability targets are almost always quoted per component and experienced per request, and that gap is where the surprise lives.
The error budget is the useful reframe. If you are allowed 43 minutes of downtime a month, that is not a failure quota, it is a spending allowance. It buys risky deploys, migrations, and the experiments you would otherwise never run. A team that ends every month with its full budget unspent is not reliable, it is too slow.
One caution on redundancy: the parallel maths assumes copies fail independently, and they usually do not. Two instances in the same rack, on the same power, behind the same load balancer, running the same bad deploy, are far closer to one instance than the arithmetic suggests.
Questions people arrive with
- How much downtime is 99.9% uptime?
- About 43 minutes a month, 8 hours 46 minutes a year, and only 1 minute 26 seconds a day. Three nines sounds close to perfect and permits most of a working day of outage every year. The calculator above converts any target into all four figures.
- What is the difference between three nines and four nines?
- A factor of ten. 99.9% allows about 8.8 hours of downtime a year and 99.99% allows about 53 minutes. Each additional nine divides the permitted downtime by ten and usually multiplies the engineering cost by considerably more than that.
- Why is my real availability lower than each service's uptime?
- Because services in series multiply. Five services at 99.9% each, all of which must work for a request to succeed, give about 99.5% overall, which is roughly 44 hours a year rather than nine. Availability is quoted per component and experienced per request, and that gap is where the surprise lives.