99.95% uptime: how much downtime is that?

99.95% uptime in exact figures

Most SLAs are measured per month, and a real month is 28 to 31 days long. The same 99.95% allows this much in calendar months and years:

Formula: downtime = (100 - 99.95) / 100 × period. The averages use a 365.2425-day year (the average Gregorian year, leap days included), a month of 1/12 of that and a quarter of 1/4. How the calculator works

What 99.95% uptime means in practice

Typical for: Managed databases and services with a hot standby, paid SaaS tiers that promise more than three nines, and APIs that other businesses build on.

• Failover has to be automatic or very fast. A manual switch to a replica that takes 20 minutes uses most of the month.
• Deploys must not cause downtime: rolling or blue-green deploys, and database migrations that do not lock tables for minutes.
• Someone has to be on call. With 21m 54.9s a month, an alert that waits until the morning breaches the SLA.

Maintenance windows

Planned downtime does not fit unless the SLA excludes it. Most teams at this level do maintenance with the service up, or agree an exclusion for windows announced in advance. One 30-minute window counted as downtime uses 137% of the monthly budget, more than the whole month allows.

Monitoring and alerting

Check at least every minute, and every 30 seconds if you can. Keep confirmation short: a few retries seconds apart, not several failed checks minutes apart.

An outage that starts just after a check can run for a whole interval before the next check sees it. This is what one such interval costs out of the 21m 54.9s a month that 99.95% allows:

Drumbeats uptime checks run as often as every 60 seconds on the Free plan and every 30 seconds on paid plans; you set the interval for each monitor. A failed check is retried twice, 5 seconds apart, before it counts as down, so a single failed request does not page you. Uptime monitoring

How one incident spends the 99.95% budget

The primary database fails over by hand

A typical failure at this level: the primary database stops answering and an engineer promotes the replica. The service is checked every 30 seconds.

The service was down for 15m 40s, 71% of the 21m 54.9s budget, which leaves 6m 14.9s for the rest of the month. There is room for one failure like this; the second breaches the SLA. Automatic failover removes the two slowest steps, which is why most teams at this level automate it.

Downtime counts from the moment the service fails, not from when a check or a person notices it, so the time it takes to detect a failure comes out of the budget as well as the time to fix it.

Who commits to 99.95%

• Google Compute Engine commits to 99.95% monthly uptime for a single VM of the memory-optimized family, in most regions (Premium Tier network). Compute Engine SLA

Checked against each provider's SLA page in October 2026. Providers change their terms; the linked page is the one that counts.

99.95% compared with nearby levels

Each extra nine cuts the downtime budget by a factor of 10. Full table from 90% to 99.999%

Questions

How many minutes of downtime is 99.95% uptime?

About 21.91 minutes a month (21m 54.9s) and 4.38 hours a year (4h 22m 58.5s). Per day it is 43.2s.

Why 99.95% and not 99.9% or 99.99%?

It sits between them: half the downtime budget of 99.9% and five times that of 99.99%. It suits a service with redundancy that still relies on people for some recovery steps.

What is three and a half nines?

Another name for 99.95%. It has three nines followed by a five, and sits between three nines (99.9%) and four nines (99.99%).