Architecture & System Design › Reliability & Resilience

Nines (99.9%, 99.99%)

How much downtime each level of availability allows.

Also known as: nines of availability, 99.9%, availability nines

Nines quantify availability as 9s in the percentage: 99.9% (three nines) allows ~43 minutes of monthly downtime; 99.99% allows ~4 minutes; 99.999% allows ~26 seconds. Each nine shrinks the error budget tenfold — and raises the engineering cost roughly tenfold with it.

99%       ≈ 7h 18m  monthly budget (two nines)
99.9%     ≈ 43m     (three nines — the common serious target)
99.99%    ≈ 4m      (four nines — redundancy everywhere, tested constantly)
99.999%   ≈ 26s     (five nines — specialised, expensive, rarely needed)

Nines only mean something scoped: which service, which requests, measured how, over what window, excluding what. Unscoped nines (“the platform is four nines”) collapse the moment anyone asks what counts as down.

The classic mistakes:

• Nines without scope or measurement. Targets nobody measures are aspirations. Define SLI, window, exclusions — then the nines follow from data.
• Chasing unneeded nines. Four nines on an internal batch dashboard wastes the budget that checkout needs. Tier by criticality; spend nines where downtime costs.
• Ignoring the window. Yearly 99.9% (~9h) tolerates a bad day; monthly 99.9% (43m) doesn’t. Windows shape behaviour — monthly/rolling windows keep urgency honest.
• Excluding everything. “Excluding planned maintenance, dependencies, acts of cloud…” eventually excludes all downtime. Exclusions need user-legitimacy, not comfort.
• Dependency nines unmultiplied. Three 99.9% dependencies in series yield 99.7% — composition degrades. Budget end-to-end from the chain, not per link.
• Error budget as decoration. Budgets nobody enforces (ship freely at zero budget) change nothing. Gate releases and risk-taking on remaining budget.
• Confusing nines with reliability. Availability is one facet; correctness, durability and latency have their own budgets. Nines measure serving, not perfection.

How to use them: scope, measure, tier by criticality, enforce via error budgets. Nines are a spending discipline for downtime — allocate where it buys the most trust per dollar.