99.99% Uptime: What Four Nines Really Requires
99.99% uptime — 'four nines' — allows 4 minutes and 23 seconds of downtime per 30-day month, and 52 minutes 36 seconds per year. That is shorter than most human response times to a page, so four nines cannot be achieved by people reacting to alerts: it requires automated failover across Availability Zones.
Per day
1.4 min
Per week
10.1 min
Per month (30d)
43.2 min
Per quarter
2.2 h
Per year
8.8 h
Join the Veojson dispatch
Get the SRE reliability starter kit with SLO templates and alerting thresholds.
One email, no spam, unsubscribe any time.
Why humans cannot be in the recovery path
A pager alert takes 1–3 minutes to fire and reach someone, another 2–5 minutes for that person to open a laptop and orient, and longer still to act. That alone exceeds the entire monthly budget. Four nines therefore means health checks that eject a failing instance in seconds, multi-AZ redundancy with automatic traffic shifting, database failover that does not need a human decision, and deployments that are automatically rolled back on an error-rate signal.
The dependency maths gets unforgiving
At this level every dependency matters. Three services in series at 99.99% each give 99.97%, which already breaches the target. Anything in your critical path that is only 99.9% caps your achievable availability at 99.9% no matter what you do — so either make it redundant, cache around it, or design the product to keep working when it is unavailable.
Measurement has to keep up
A monitoring interval of five minutes cannot even detect a four-minute outage. Check at least once per minute from multiple external regions, and prefer request-based measurement — the percentage of successful requests — over a binary up/down probe, because partial failures are the ones that quietly eat a four-nines budget.
99.99% uptime allowed downtime
| Period | Allowed downtime |
|---|---|
| Per day | 8.6s |
| Per week | 1m 1s |
| Per 30-day month | 4m 23s |
| Per quarter | 13m 9s |
| Per year | 52m 36s |
What each level typically requires
| Target | Monthly downtime | Typical architecture |
|---|---|---|
| 99.9% | 43m 50s | Single region, health checks, fast rollback |
| 99.95% | 21m 55s | Multi-AZ, automated deploys, tight monitoring |
| 99.99% | 4m 23s | Multi-AZ active-active, automatic failover |
| 99.999% | 26s | Multi-region active-active, no human in the loop |