What Uptime Numbers Actually Allow
A year is 8,760 hours. Three nines, 99.9 percent, leaves one tenth of one percent for everything that can go wrong, and one tenth of one percent of 8,760 is 8.76 hours. That is what uptime numbers really are. They read like a promise and they work like a budget, and this budget allows your site to be gone for about nine hours a year.
Nine hours, and you do not get to pick them
Those nine hours do not arrive as a drip. Nobody loses four seconds a day. Downtime comes in a lump, because the things that cause it are lumpy. A drive fails. An update goes bad. A switch dies in a rack in a building you have never seen. You can spend the entire annual allowance in one Saturday morning and the provider is still comfortably inside the number they sold you.
The hours are not weighted either. A florist who loses nine hours spread across three quiet Tuesdays in February is mildly annoyed. The same florist who loses four hours on the morning of February 13 has lost a real slice of the year. The agreement counts those the same way, because an hour is an hour to a monitoring system. Customers do not experience an average.
This is worth saying plainly when somebody is comparing plans. A page selling web hosting for small business that leads with 99.9 percent is telling the truth about the arithmetic and almost nothing about the risk, because the risk is timing, and the percentage has no opinion about timing.
The contract counts hours. Your customers only ever count the one they showed up in.
The meter resets every month
Look at where the meter sits. Most guarantees measure by calendar month, not by year. Three nines across a 30-day month is 43 minutes, which sounds much tighter than nine hours, and per month it is. Then the meter resets on the first. Forty minutes down in March and forty minutes down in April are two compliant months, and nothing is owed.
There is usually a floor as well. Plenty of agreements only start counting once an outage runs continuously past a threshold, commonly five or fifteen minutes, and count in whole increments after that. A site that goes dark for three minutes every afternoon during a backup window can be unreachable at a predictable hour every single day and still record zero downtime.
The measuring is theirs, too. Their monitoring, from their vantage point, usually inside or right next to their own network. A customer of yours on a phone in Cheektowaga getting a timeout is not a data point. If you want an outage counted, you generally have to prove it happened, using their logs, on their terms.
What the uptime numbers do not cover
The exclusions are where the number gets most of its confidence. Scheduled maintenance is almost always outside the count, and so is emergency maintenance. So is anything the provider attributes to a denial of service attack, a third-party network, a routing problem past their edge, or weather. So is anything caused by your own code, your own plugin, your own theme, or your account going over a resource limit.
That last one covers more ground than people expect. If the server rolls PHP forward and the contact form starts throwing a fatal error, the server is up. If a plugin update takes the checkout down at 2am, the server is up. If traffic from a good week trips the memory ceiling on your plan and visitors start getting a 500, the server is up and you were the cause. All of that sits on your side of the line.
Subtract the exclusions and what remains covered is roughly the machine and the wire running into it. That is the part that fails least often and gets fixed fastest. The guarantee is strongest in exactly the place you needed it least.
The number covers the hardware. Most bad mornings are not the hardware.
The credit is not the loss
Then there is the remedy. In the typical case, a qualifying outage pays a service credit, some percentage of that month's fee per hour of downtime, capped at the month's fee. Put real money in it. Hosting at 20 dollars a month with a 10 percent credit per qualifying hour gives you two dollars off next month for an hour offline. If the site were dark for the entire month, the most you could recover is the 20 dollars you paid.
You also have to ask. Credits are almost never automatic. There is a claim window, often thirty days, and you are expected to submit dates, times and a description in writing. Nobody running a two-person shop is filing a downtime claim for two dollars, and that is part of why the guarantee is cheap to offer.
The credit is also credit. No money leaves their building. It is applied against future service from the same company, so the compensation for failing you is that you keep buying from them. The actual cost of the outage, the Saturday the bakery took no online orders, the wedding job that went to whoever answered first, appears nowhere in the agreement, because there is no way to write it in a form a provider would sign.
Down is not the only way to be broken
Being up is not the same as working. Uptime monitoring asks for a page and checks that something came back. A site taking twelve seconds to load returns a perfectly healthy 200 and counts as available. Customers do not wait twelve seconds. There is no credit for slow, because almost no agreement bothers to define slow.
SSL is the clean example. A certificate renewal fails quietly at three in the morning, and by breakfast every visitor gets a full page browser warning saying the site is not safe. The server is up. The pages are serving. The availability graph is a solid green line and your traffic is at zero.
Email is worse, because it is usually measured separately or not measured at all. Quotes, orders and invoices move by email in most local businesses. A mail queue backing up, or your address landing in spam folders for a week after a neighbor on a shared IP misbehaved, does not touch the site's availability figure, and it can still be the most expensive week of your year.
Domains pull the same trick from another direction. A card on file expires, the renewal fails, the domain lapses, and the business is off the internet completely. None of that is downtime in the hosting sense. The server never missed a beat. That is why hosting and domain support is one job and not two, and why the useful question is who is watching the expiry dates, not what percentage is printed on the plan.
What is actually worth buying
So the guarantee is not a useful thing to shop on. Three nines allows 8.76 hours a year, four nines allows about 53 minutes, and the two companies quoting them are often running similar hardware in similar buildings. The gap between the printed numbers is much smaller than the gap between what happens next.
What separates hosts is the hour after it breaks. Whether a person answers. Whether that person can find your account and read a log instead of telling you to clear your cache. Whether there is a backup from last night, how long a restore truly takes, and whether anyone calls you back when it is done. None of that carries a percentage, which is exactly why it is not on the pricing page.
Knowing before your customer does matters just as much. Most small businesses find out their site is down because somebody phoned to mention it. Monitoring you control is worth more than a paragraph of legal text promising two dollars, because it turns a five-hour outage into a twenty-minute one.
Three nines is not a lie, and the companies quoting it are not crooks. It is an honest description of a budget, written by lawyers, measured with their instruments and settled in their currency. Read it as what it is. Here is roughly how much we expect to fail. Here is the small amount we will refund if you ask correctly and on time. Here is the long list of things we are not counting. Then ask the question the percentage cannot answer, which is what happens on the morning the nine hours turn out to be yours.
More from the blog
- What Unlimited Hosting Really Means
- The First Hour When Your Site Goes Down
- Why Your Business Email Goes to Spam
All posts are on the blog. For the step by step versions, see the domain and hosting guides.