All articles
7 min read

What 99.9% uptime actually allows

Every extra nine costs roughly ten times more. Here is the downtime each one permits, and why the number on the SLA is rarely the number that matters.

“Three nines” sounds like a guarantee of near-perfection. It permits about 43 minutes of downtime a month, which is long enough for customers to notice, tweet about it, and open support tickets. The arithmetic is worth internalising before you promise anything.

TargetPer dayPer monthPer year
99%14m 24s7h 12m3d 15h
99.9%1m 26s43m 12s8h 45m
99.95%43s21m 36s4h 22m
99.99%8.6s4m 19s52m 34s
99.999%0.86s26s5m 15s

Monthly figures use a 30-day month. The pattern is that each additional nine divides the allowance by ten, while the engineering effort to achieve it multiplies by considerably more than ten.

Four nines is a different kind of system

52 minutes of downtime per year is less than one careless deploy. At that level you cannot take the service down to migrate a database, you cannot have a single load balancer, and you cannot wait for a human to wake up and read an alert. Four nines is not a harder version of three nines — it is a different architecture, with the on-call rota and the cost that implies.

The measurement matters more than the target

An SLA number means nothing without its definition, and the definition is where the argument happens. Three questions decide what the figure really promises:

  • How often is it checked? A one-minute interval can miss a 45-second outage entirely. If the SLA is measured by five-minute polls, a brief but total failure may never appear in the number at all.
  • From where? A service reachable from Frankfurt and unreachable from Singapore is up by one measurement and down by another. Single-region measurement quietly flatters the number.
  • What counts as down? A 200 response containing an error page is up by any naive check. Scheduled maintenance is usually excluded, which is exactly why the excluded window tends to grow.

This is why uptime should be time-weighted rather than a simple ratio of successful checks. If you change a monitor from five-minute to one-minute checks, a ratio of checks silently rewrites your history; weighting each result by the time it represents does not.

Picking a number you can keep

For most B2B software, 99.9% is the honest commitment: it is achievable without a follow-the-sun rota, and it leaves room for the occasional bad deploy. Promise it only if you measure it — publishing a target you do not track is a claim, not a commitment, and the first serious outage is when customers discover which one it was.

Monitoring that tells you before your customers do

Websites, APIs, ports, DNS, certificates and cron jobs — checked from several regions and confirmed across them before you get woken. Free plan includes 5 monitors.

Start monitoring free