Availability & Failover

Availability is usually quoted in nines, and the arithmetic is worth knowing cold because interviewers ask for it directly.

What the nines mean per year: 99% is about 3.65 days of downtime. 99.9% ("three nines") is about 8.8 hours. 99.99% is about 53 minutes. 99.999% is about 5 minutes. Each extra nine costs roughly an order of magnitude more effort, which is why "five nines" is a claim to interrogate rather than a target to assume.

Components in sequence multiply. If a request must pass through two services that are each 99.9% available, the path is 0.999 x 0.999 = 99.8%. Dependencies make you less available, and a long chain of healthy-looking services can add up to a bad number. This is the single most useful availability calculation in an interview.

Components in parallel compound the other way. Two redundant 99.9% replicas give 1 - (0.001)² = 99.9999%, because both must fail simultaneously. Redundancy is how you buy nines; dependencies are how you spend them.

Failover shapes. Active-passive keeps a standby that takes over on failure — simpler, and you pay for idle capacity while the failover window is real downtime. Active-active serves from both, so there is no cutover and you get load distribution, but now you own write conflicts and both sides must be able to carry full load alone, or failover just moves the outage.

The gap people miss: failover machinery is itself a system that can fail, and it is exercised rarely. An untested failover path is a hypothesis, not a guarantee.