Availability (the nines)

Availability is uptime expressed as a target rather than as a measurement, and it is spoken in nines: 99.9% is three nines, 99.99% is four, 99.999% is five. Each additional nine divides the permitted downtime by ten, which is why the ladder gets expensive so quickly — the difference between three and four nines is the difference between an outage you can debug during business hours and one you must detect, page for and resolve inside five minutes.

The table below is the whole of the arithmetic. It is worth reading before agreeing to a number in a contract, because the figures are smaller than most people picture: three nines allows 8h 45m 36s a year of the kind of outage that would end up in a postmortem, and five nines allows 5m 15s — roughly the time it takes to notice.

Availability can also be calculated from reliability metrics: Availability = MTBF / (MTBF + MTTR). A service that runs 238 hours between failures on average and takes 2 hours to recover is available 238 / 240 = 99.17% of the time. The formula makes the trade visible: halving the recovery time buys exactly as much as doubling the time between failures, and recovery is usually the cheaper of the two to improve.

Two caveats change what these numbers mean in practice. First, availability is almost never measured against your whole system — it is measured against whatever the monitoring checks, so the target is only as honest as the check behind it. Second, a target that cannot be measured at the resolution it promises is decorative: promising five nines while checking every five minutes means a single failed check has already spent nearly a year's budget.