Availability Math & the Nines

Availability, in minutes you can feel

Availability is the fraction of time (or of valid requests) a service is up and serving correctly. People quote it in "nines," and the single most useful senior habit is to translate nines into minutes of allowed downtime per month, because that is what an on-call rotation actually feels.

The math: allowed downtime = (1 - availability) x window. For a 30-day month (43,200 minutes):

Nines     Availability   Downtime / month   Downtime / year
two        99%            ~7.2 hours          ~3.65 days
three      99.9%          ~43.8 minutes       ~8.76 hours
four       99.99%         ~4.4 minutes        ~52.6 minutes
five       99.999%        ~26 seconds         ~5.26 minutes

Notice the leap between each row. Going from 99.9% to 99.99% shrinks your monthly downtime budget from 43.8 minutes to 4.4 minutes. That is not "a bit better," it is a 10x reduction in the failure you are allowed, and every added nine costs roughly 10x more to achieve. The reason: the cheap failures (a bad deploy, a full disk) are gone by three nines, so the next nine forces you to attack rare, expensive causes: multi-AZ redundancy, automated failover measured in seconds, eliminating every manual step from recovery, and testing failure paths constantly. Human response time alone (someone gets paged, opens a laptop, diagnoses) blows a five-nines budget, so five nines effectively means no human in the recovery loop.

Dependencies combine

Serial dependencies multiply. If your checkout calls auth, inventory, and payments in series and each is 99.9% available, your ceiling is 0.999^3 = 99.7%, worse than any single component. More hops means a lower ceiling. You cannot be more available than the product of everything you synchronously depend on.

Redundancy adds availability. Two independent replicas of a 99% component, where either can serve, fail only when both fail: 1 - (1 - 0.99)^2 = 99.99%. Parallel combines as 1 - (1 - a)^n. This is why the fix for a shaky dependency is often a second independent copy, not a more reliable single copy.

Interview nuance: three different numbers

Interviewers probe whether you distinguish three different numbers. Measured availability is what your telemetry actually observed last month. The SLA is the external contractual promise with financial penalties (service credits) if you miss it. The SLO is your stricter internal target, deliberately tighter than the SLA so you get warning before you breach the contract. A team runs to a 99.95% SLO to safely honor a 99.9% SLA.

Common wrong turn: chasing five nines everywhere. If your database ceiling is 99.9% and a feature earns 20 dollars a minute of downtime saved, spending a quarter's engineering to add a nine it can never reach is malpractice. Match the target to revenue impact and to the dependency ceiling.

Recap: convert nines to downtime minutes, remember serial dependencies multiply (lowering the ceiling) while redundancy combines as 1 - (1-a)^n, each nine costs about 10x more, and keep measured, SLA, and SLO as three separate numbers.