Uptime

Uptime describes how reliably a service can be reached. This can refer to a website, an app, an online game, or a payment system. To measure it, a period of time is observed, for example a month, and it is counted how long the service worked normally during that time. This proportion is stated as a percentage. 99.9 percent uptime means: out of a thousand hours, the service was disrupted for one hour. The counterpart is called downtime, meaning the time during which nothing works.

Why the decimal places are worth a lot of money

The numbers look similar but mean very different things. 99 percent uptime allows for around 3.7 days of downtime per year. 99.9 percent leaves only about 8.8 hours. At 99.99 percent, it’s 53 minutes for the entire year. Every additional nine costs significantly more effort, because it requires additional technology and personnel. Experts therefore speak of “three nines” or “four nines.”

For companies, downtime is directly costly. An online shop sells nothing during an outage. A bank cannot process transfers. With a cloud provider, thousands of customer companies also depend on it, and all of them come to a standstill at the same time. Large outages therefore regularly make it into business news.

Providers often give their customers a written commitment regarding uptime. This document is called a Service Level Agreement, or SLA for short. If the promised uptime is not met, the customer usually receives a credit. The fine print is important here: scheduled maintenance windows often do not count as downtime.

How providers prevent outages

The most important tool is called redundancy. This means that every important component exists in duplicate. If one server fails, a second one immediately takes over. The same applies to power supply, network lines, and cooling. Large providers even distribute their systems across multiple data centers in different regions.

A comparison helps: an airplane has multiple engines so that a defect doesn’t immediately lead to a crash. Data centers are built the same way. A program continuously monitors all components and automatically redirects requests as soon as something stops responding. This switchover is called failover.

Uptime is measured through monitoring. Automated test programs call up the service every minute from various locations around the world. If it doesn’t respond or responds too slowly, this is logged and an alarm is triggered. A common misconception is to confuse uptime with speed. A service can be reachable and yet painfully slow — the uptime then formally remains high.

Uptime in cloud services and AI offerings

Uptime is most visible on status pages. Almost every major provider runs such a page, for example status.openai.com. It shows in traffic-light colors which parts are currently running and where there are problems. After a major outage, many companies additionally publish a report on the cause.

With AI services, the topic is particularly sensitive. Many companies build their own products on top of a provider’s programming interface, meaning an access point through which one piece of software uses another. If this access point fails, all products built on top of it also come to a standstill. Because graphics cards for AI are scarce and expensive, overloads are more common here than with classic web services.

You encounter uptime in everyday life even without the word being mentioned. When a streaming service crashes on the evening of a finale, that’s downtime. When a bank blocks its app at night for maintenance, that’s scheduled downtime. And when a game server is overloaded at the launch of a new title, users and the press discuss exactly this metric.