Five Nines Uptime: What 99.999% Really Costs

Five nines is shorthand for 99.999% availability, which allows roughly 5 minutes and 15 seconds of downtime across a full year. The number is easy to promise and surprisingly hard to prove. Most teams that claim it have never checked whether their own monitoring can resolve an outage that short.

What does five nines mean?

The phrase counts the nines in an availability percentage. Three nines is 99.9%, four nines is 99.99%, and the fifth nine takes you to 99.999%. SUSE's definition page puts the annual budget at about 5.26 minutes of total downtime, planned or unplanned.

That last clause is the one people skip. Planned maintenance spends the same budget as an outage, so a single 10-minute deploy window burns two years of allowance in one evening.

How much downtime does each nine allow?

Each additional nine cuts the allowance by a factor of ten. Splunk's availability primer publishes the same table, and it is worth reading as a repair-time budget rather than as a scoreboard.

Availability	Per year	Per month	Per week
99%	3d 15h 36m	7h 18m	1h 41m
99.9%	8h 45m	43m 28s	10m 5s
99.95%	4h 22m	21m 44s	5m 2s
99.99%	52m 35s	4m 21s	1m 0s
99.999%	5m 15s	26s	6s

The jump from the fourth nine to the fifth is the expensive one. You go from four and a half minutes a month, which a human can respond to, down to 26 seconds, which nobody can. Everything at that tier has to detect and fail over without waking anyone up.

Why can't an external check prove five nines?

Because the allowance is smaller than the measurement resolution. A check running every 30 seconds samples availability about 86,400 times a month. Each sample is a point in time, not an interval, so the shortest outage it can place is bounded by the gap between two checks.

Work the arithmetic through. The monthly budget at this tier is 26 seconds. A 30-second interval cannot tell zero seconds of downtime apart from 29, and confirmation adds more delay on top. In our testing a failing endpoint checked every 30 seconds with two-region confirmation moved from SUSPECT to DOWN in 60 to 90 seconds. That detection floor is already triple the entire monthly budget.

Four nines behaves differently. A 4 minute 21 second monthly allowance sits comfortably above a 90-second detection floor, so an outside-in check can measure it honestly. That is why a 99.999 uptime claim usually rests on internal instrumentation rather than on a probe outside the network. Google's SRE book argues the same thing from the other side: pick a target you can actually observe.

What to measure instead

Set an objective you can defend with evidence, then watch the error budget burn down. Three or four nines on one named endpoint beats an unmeasurable claim about a whole system. Give the check an explicit response-time assertion. Slow should register as degraded rather than fine:

curl -X POST https://uptimeroom.com/api/v1/monitors \
  -H "Authorization: Bearer urm_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "checkout",
    "type": "http",
    "target": "https://example.com/health",
    "intervalSec": 30,
    "regions": ["los-angeles", "london", "singapore"],
    "expectedStatus": [200],
    "responseTimeThresholdMs": 1000
  }'

Run that monitor from three of the six probe regions available as of September 2026. We prefer three over two, because two regions give you agreement or deadlock and nothing in between. Egress addresses for each region are published on the probe IPs page if you need to allowlist them.

Common mistakes

Averaging availability across every monitor. Ten dependencies at 99.9% each do not add up to a 99.9% service. If one request touches all ten in sequence, compound availability lands near 99%. Multiply the numbers, do not average them.

Counting one failed check as downtime. A single probe failure is a blip, and blips inflate your outage minutes badly at this scale. Confirm from a second region before the clock starts, which is what cross-region quorum confirmation exists to do.

Publishing a mean response time beside the availability figure. The mean hides the tail where the unhappy users are. Keep p50, p95 and p99, the same discipline the four golden signals apply to latency.

Treat five nines as a budgeting exercise and not as a badge. Work out what an hour of downtime actually costs you, weigh it against the redundancy bill, and most services land at 99.9% or 99.95% once payments are excluded. Then measure that target honestly with a 30-second interval and multi-region confirmation. Both are included on every paid plan listed on the pricing page, alongside a year or more of history.