What One More Nine Actually Costs You in Minutes
Subtitle: A nine is a downtime budget.
- Left column - Budget per year:
- 99% = 3 days 15 hours
- 99.9% = 8 hours 45 min
- 99.99% = 52 min 34 sec
- 99.999% = 5 min 15 sec
- Less than one careless deploy
- Right column - What a nine costs:
- Each nine cuts the budget by 10
- Engineering cost rises far faster
- Not 'can we be more available'
- Ask 'what is this nine worth'
- Buy a nine only where it pays
- Nines multiply down the chain:
- Your 99.9% is only a ceiling
- 4 sync deps at 99.9% drag it down
- Composition alone costs 0.7 nines
- Remedy: timeouts, fallbacks, async
- Simple difference:
- 0.999 ^ 5 = 99.5% = 2.3 nines
- 8h 45m budget becomes 1d 19h 43m
- Quoted about the wrong thing:
- Per month 99.9% = 43 minutes
- Per year 99.9% = 8h 45m
- Same SLA, 12x stricter window
- It is an average over requests
- One dead region hides in green
Sticky note - Honest test:
- Compute your nines from raw request logs, per region. Compare that with the dashboard headline.