Availability from MTBF and MTTR

Availability is uptime divided by total time, which written in reliability terms is $\mathit{MTBF}/(\mathit{MTBF} + \mathit{MTTR})$. A thousand hours between failures and four hours to fix gives 99.6%, which sounds excellent and is about 35 hours of downtime a year. The nines are a compressed scale and they hide how much work each one costs.

Read it as a ratio and the design lever appears. Availability depends only on how MTBF compares to MTTR, so three nines means uptime is 999 times repair time, and four nines means 9999 times. There are two ways to get there: make the thing last ten times longer, or fix it ten times faster. Extending MTBF means better hardware and it gets expensive fast. Cutting MTTR means monitoring, spares on the shelf, a runbook and a tested failover, and it is almost always the cheaper nine. A system that fails weekly but recovers in four seconds beats one that fails yearly and takes a day.

The classic mistake is multiplying availability figures without asking what they are attached to. Chain four components at 99.9% and the system is $0.999^4 = 99.6\%$, so buying four excellent parts got you a mediocre whole. Note also that this is inherent availability: it counts repair time only, and excludes scheduled maintenance, waiting for a part, and the two hours before anyone noticed. Operational availability is always worse, and the gap between the two is where most missed SLA targets actually live.