99.999% Availability

The Downtime Math

 99.999% → 0.001% failure budget

 per YEAR    ≈ 5.26 minutes down    (525,600 × 0.001%)
 per MONTH   ≈ 26 seconds
 per WEEK    ≈ 6 seconds

Five minutes a year. One bad deploy consumes a decade of budget. At this tier the question stops being “how do we recover fast?” and becomes “how do we never go down at all?” — which changes architecture fundamentally.

What Five Nines Physically Requires

No single failure of any kind may be user-visible:

 - NO in-place changes to serving paths: canary or nothing;
   rollbacks measured in seconds
 - active-active across regions: traffic shifts on health signal,
   not human decision (humans are slower than the budget)
 - data layer: consensus replication (quorum writes) with regions
   as failure units — sync cost paid on every write
 - zero shared dependencies: each region fully serves alone
   (auth, config, feature flags replicated locally)
 - network redundancy: multiple providers/anycast; BGP failover
   rehearsed, not assumed
 - clients engineered for it: retries with jitter, stale-tolerant
   reads, offline modes — because SOME sub-second blips remain

The client point deserves emphasis: at five-nines budgets, user-perceived availability includes DNS resolvers, home Wi-Fi, and mobile networks — none under your control. Server-side five nines cannot guarantee perceived five nines; products engineer graceful client behavior instead.

Who Actually Achieves It

Operator	Why they can
Telecom switches / PSTN core	Decades-hardened hardware + software dual-rail
Payment networks (card rails)	Purpose-built, massively redundant, slow to change
Airline/defense control systems	Certification regimes; cost no object
Hyperscaler internal fabrics	Custom everything; fleets of PhDs

Notice who is absent: typical SaaS products and consumer apps. Public claims of five nines usually measure components, marketing arithmetic, or nothing.

The Economics Are Brutal

Each nine multiplies cost roughly 2–10x:

 99.9%   → 99.99% : multiples (multi-region readiness, progressive delivery)
 99.99%  → 99.999%: another order-of-magnitude jump
                    AND velocity tax: every change passes heavy gates,
                    feature speed drops measurably
 
 five nines is purchased with AGILITY — the currency most
 growing products need more than minutes of yearly downtime

The rational exception set is tiny: money movement, safety systems, contractual telecom. Everything else buying five nines is buying status, not uptime.

Interview Framing

When candidates volunteer five niles unprompted, experienced interviewers hear miscalibration. The strong pattern stays consequence-driven: “five nines means active-active multi-region with consensus writes and change freezes — I’d only put payment capture near that tier, and even there four nines with instant rollback is usually the honest engineering choice.” Knowing when NOT to buy the top tier is the seniority signal this number exists to test.