One shard fails for 30 minutes. Is today 99.58%, 97.92% or 100%?
Count failed requests, bad minutes under 95%, or full outages only. Write the rule down first, and count errors, too-slow and wrong answers.
4.32
minutes of downtime a month at 99.99%
Each nine divides the budget by ten, from 43.2 minutes at 99.9%. Four nines is spent before an alert fires and an engineer starts looking.
Four hard dependencies make checkout worse than its weakest link.
Every box must be up, so multiply. Near 100%, the downtimes simply add up, to about 91 minutes a month.
Two 99.9% replicas: six nines on paper, three and a half live.
Independent, they’d fail together 32 seconds a year. A shared AZ and config push cost 20 minutes a month, so the pair is ≈ 99.95%.
99.90% → 99.98%: which lever?
Bad: Fail less often
Needs one incident every 135 days, not monthly: 4.5× fewer. Far harder.
Good: Recover faster
Alert in 2, diagnose in 3, auto-rollback in 5: MTTR 45 → 10 minutes.
EC2’s SLA says 99.99%. A bad month delivers 99.9%: 43 minutes with every AZ unreachable.
What AWS pays
A $2,000 credit on a $20,000 bill. At a ₹2 crore-an-hour peak, you lost ≈ ₹1.4 crore.
Starting targets per journey, priced by what failure costs
- Payments: 99.99% or higher, because money moves
- Sale-day checkout: 99.95–99.99% for the event
- Search: 99.9–99.95%, degrades to cached results
- Recommendations: 99.5–99.9%, soft, so hide the widget
Remember
Measure availability per journey: valid requests that succeed in time. Each nine is 10× less downtime. Hard dependencies multiply; replicas help only without shared fate. The cheapest nine is faster recovery.