99.9% Uptime

Also known as high availability for network services, though the definition of "acceptable service" varies. Did you get that contract in writing? Another problem is that chasing additional nines can become very expensive in terms of system design and support—can future hires maintain this thing?—and even then what happens when your upstream DNS provider goes kablooie for a bit too long?

Assuming roughly 31,536,000 seconds in a year, your allowed downtimes for various percentages are, depending on the whims of floating point math and possible CPU (rare) or compiler (also rare) bugs,

    $ cat up
    #!/usr/bin/env perl
    my $total  = 86400 * 365;
    printf "%9.9f %d\n", 0.09, $total - sprintf "%.f", $total * 0.09;
    my $nines  = "0.9";
    my $uptime = ~0;
    do {
        $uptime = $total - sprintf "%.f", $total * $nines;
        printf "%9.9f %d\n", $nines, $uptime;
        $nines .= "9";
    } while $uptime > 0;
    $ perl up
    0.090000000 28697760
    0.900000000 3153600
    0.990000000 315360
    0.999000000 31536
    0.999900000 3154
    0.999990000 315
    0.999999000 32
    0.999999900 3
    0.999999990 0

We might agree that a 9% uptime is pretty bad—over 47 weeks of allowed downtime in a year!—though there are apparently insects that only spawn every 17 years or so, so worse uptime values are not unknown. Dismal economic prospects for the lot of them! Otherwise as the nines increase one quickly gets to a point where really no downtime is allowed, so if a system or network router burps for a second or two, well that screwed your numbers for the year. The usual 99.9% allows for nearly nine hours of outage, which isn't that much over a whole year, especially when there are humans meddling with the system, or if there's old or faulty gear installed. It may take someone wandering around Belgium in the wee hours of the night several hours to find a replacement motherboard for a database server, for example.

    $ sumtime 31536
    8h45m36s

Some sites "cheat" and do not count downtime during scheduled downtime, only if the downtime strays outside said scheduled windows. This is probably a good, pragmatic compromise, as it limits the need for more expensive or complicated solutions that those higher uptime numbers really do need.

On the other hand, the notion that systems need to be 100% available is slightly silly, given that humans only have a daily uptime of around 66.6% (assuming eight hours of sleep) and less than that if you include time to burn off stress and so on and so forth. Does the uptime number need to go up?