Four nines is a budget, not a grade

An availability target sounds like a quality grade. 99.9% feels good, and 99.99% feels better. It is really a budget of downtime, and the budget is smaller than most people expect.

At 99.9%, you get about 8.76 hours of downtime a year, or 43.2 minutes in a 30-day month. At 99.99%, you get 52.56 minutes a year, or 4.32 minutes a month. Each extra nine cuts the allowance by a factor of ten.

Downtime allowed at each target, for a 365-day year and a 30-day month.

Everything comes out of that one allowance: planned maintenance, a bad deployment, a dependency that fails at 2 a.m., and the incident nobody has had yet. At 99.99%, a single 45-minute maintenance window spends about 86% of the year's budget before anything has actually broken.

Dependencies make it tighter. Two components in series, each at 99.9% with independent failures, give you about 99.8% end to end. Redundancy helps, but only as far as the failures are truly independent. Shared power, a shared network path, shared configuration or one deployment pipeline can take both sides down together.

Before you commit to a number, run a short test:

1. Convert the target into minutes per month.
2. List everything that will spend those minutes: maintenance, deployments, dependencies and unplanned incidents.
3. Decide who owns the budget and what happens when it runs out.

If the honest list does not fit the budget, change the target or the design before the number goes into a contract, a dashboard or a slide.

The full guide covers time-based versus request-based availability, the arithmetic behind each figure, and what Google's SRE book says about error budgets: https://365technical.com/guides/availability-nines-downtime/

I write about IT infrastructure, operations and leading technical teams. My book Modern IT Infrastructure and Operations goes deeper on resilience and disaster recovery: https://link.amazon/B0bfFEK1w