Every provider publishes an availability figure and it is one of the least useful numbers in the comparison. Not because it is dishonest, but because it answers a different question from the one people think they are asking.
The percentages, in minutes
Availability figures are usually quoted per month. Converted into downtime you would be entitled to complain about:
- 99% — a little over 7 hours a month.
- 99.9% — roughly 43 minutes a month.
- 99.95% — roughly 21 minutes a month.
- 99.99% — roughly 4 minutes a month.
- 99.999% — under 30 seconds a month.
Those are arithmetic conversions of the percentage, not claims about any particular provider.
An SLA is a refund policy
This is the part that changes how you read it. A service level agreement does not promise the service will be up. It states what the provider owes you if it was not — and what it owes is almost always a service credit against future invoices, capped at some fraction of what you spent on the affected service.
Three consequences follow:
- The compensation is proportional to your spend, not to your loss. An outage that costs a business a day of orders is compensated with a percentage of a server bill.
- Credits are usually claim-based. If nobody files, nobody is paid.
- The measured window is per service, per region, and typically excludes maintenance you were notified about, anything caused by your own configuration, and force majeure.
Single instances usually carry a weaker promise
The headline figure normally applies to a service deployed across multiple availability zones. A single virtual machine in a single zone is generally covered by a lower commitment, because the provider is explicitly not promising that one physical host will not fail.
That is not a trap. It is the platform telling you where its guarantee ends and your architecture begins.
Where real availability comes from
Ranked by how much they actually move the number, for most workloads:
- Your own deployment process. Far more outages come from a bad release than from a hypervisor failure.
- Backups you have restored from. Recovery time is availability. An untested backup is a hope.
- Removing single points of failure you control — one database, one certificate that expires, one cron job nobody documented.
- Health checks and automatic replacement, which is what a load balancer gives you beyond distributing traffic.
- Multiple zones, which is where the platform’s own guarantee starts to apply.
Multi-region and multi-provider come after all of that, and cost considerably more than they usually return. A multi-provider strategy is worth doing deliberately or not at all.
How to read an SLA before you commit
- Find the definition of “unavailable”. Some providers count only total inability to connect, not degraded performance.
- Find the measurement interval. Five-minute intervals and one-minute intervals produce very different numbers from the same incident.
- Find the exclusions list. It is usually longer than the guarantee.
- Find the claim window. Thirty days is common, and it is easy to miss.
Where to go next
Questions
Frequently asked
Does a 99.99% SLA mean my site will not go down?
No. It defines what the provider owes you if their service is unavailable beyond the stated threshold, normally a credit against future bills. It is a commercial term, not an engineering guarantee about your application.
Do SLA credits cover lost revenue?
Essentially never. Credits are capped at a share of what you paid for the affected service in the affected period, and are unrelated to what the downtime cost your business.
Does a single virtual machine get the headline SLA?
Usually not. Headline availability figures typically apply to deployments spread across multiple availability zones; single-instance commitments are normally lower and stated separately.