What 99.9% actually means
Every provider quotes a percentage. Converted into time, the differences stop being abstract.
The table
- 99%
- 7 hours 18 minutes of downtime per month. Nearly four days a year.
- 99.9% — three nines
- 43 minutes a month, 8 hours 46 minutes a year.
- 99.95%
- 21 minutes a month.
- 99.99% — four nines
- 4 minutes 23 seconds a month.
- 99.999% — five nines
- 26 seconds a month. A single reboot blows the annual budget.
That last row is worth sitting with. Five nines means that in a whole year you may be unavailable for five minutes and fifteen seconds, including every maintenance, every kernel upgrade and every failed disk. Anyone quoting it for a single server is quoting a number they cannot keep.
What the percentage does not tell you
- What counts as downtime. Total loss only, or degraded service too? Measured from when, by whom?
- What is excluded. Announced maintenance, emergency work, attacks, your own software. The exclusions decide the number more than the number does.
- How it is measured. A provider measuring from its own network sees a very different figure from one measuring from outside.
- What happens when it is missed. An SLA with no claim process is decoration — see claiming service credits.
- One long outage or many short ones. Eight hours once a year and forty minutes a month are the same percentage and completely different experiences.
The layers add up
Your availability is the product of everything underneath: power, cooling, network, hardware, operating system, application. A 99.99% network under a 99% application gives you 99%. The weakest layer decides, and it is usually not the one being advertised.
Which is why the honest answer for most businesses is in high availability, honestly: one well-built machine with fast hardware replacement, rather than a cluster nobody has tested.
What we say
Our commitment is published in full in the SLA, with the exclusions in the same document rather than a separate page. We would rather be held to a number we keep than advertise one we do not.
In practice, service-affecting maintenance is rare here by design: capacity goes in before it is needed, hardware is replaced on a schedule rather than after it fails, and there is more than one of everything that carries traffic. See maintenance and notice.
Measure it yourself
Do not take anyone's word for it, ours included. Run an external check against your own service from somewhere that is not your server, record it, and you will have a year of real data instead of a percentage from a brochure. See monitoring your server.
Still stuck? Mail support@novogara.com — an engineer answers, at any hour. Back to the knowledge base