What the "nines" really mean
Availability (uptime) is the share of time in which a service works. Allowed downtime is simple to calculate: allowed downtime = (100 % − availability) × length of the period. On paper the difference between 99.9 % and 99.99 % is a single nine; in practice it is ten times less allowed downtime.
| Availability | Per day | Per month (30 days) | Per year (365 days) |
|---|---|---|---|
| 99 % | 14 min 24 s | 7 h 12 min | 3 d 15 h 36 min |
| 99.5 % | 7 min 12 s | 3 h 36 min | 1 d 19 h 48 min |
| 99.9 % | 1 min 26 s | 43 min 12 s | 8 h 45 min 36 s |
| 99.95 % | 43 s | 21 min 36 s | 4 h 22 min 48 s |
| 99.99 % | 8.6 s | 4 min 19 s | 52 min 34 s |
| 99.999 % | 0.86 s | 26 s | 5 min 15 s |
Most SLAs measure availability per calendar month, so the allowed downtime in February is shorter than in January. The yearly figure is useful for comparison, but the monthly one is what counts in the contract. Calculate exact values for any percentage and period with the uptime calculator on this page.
SLA vs SLO vs SLI: three acronyms, three different things
The terms are often mixed up, but they describe different levels. The Google Site Reliability Engineering book separates them as follows:
- SLI (Service Level Indicator) is a measured value, for example the share of successful requests, the share of minutes in which the service responds, or response time.
- SLO (Service Level Objective) is the target for that indicator, for example 99.95 % successful requests in a calendar month. It is an internal goal of the team.
- SLA (Service Level Agreement) is the contract with the customer that contains the promised level and the consequences of missing it, most often a service credit.
An internal SLO is therefore usually stricter than the SLA promised to customers. SRE calls the gap between the target and the actual state the error budget: while there is enough of it, the team can ship changes; when it runs out, stability takes priority.
What an SLA usually does not count
The percentage in the contract is only half of the information. The other half is the definition of downtime and the list of exclusions. When reading an SLA, check in particular:
- Planned maintenance. Many providers exclude maintenance windows announced in advance. For users the service is still unavailable, so check how long and how frequent the windows are and whether they fall outside your business hours.
- The definition of downtime. Does only a complete outage count, or a partial one too? For region-level EC2 availability, for example, AWS counts downtime only when all your running instances in two or more availability zones lose external connectivity at the same time.
- The measurement window. Monthly, quarterly or yearly. With a yearly window, a single long outage can be "diluted" over the whole year.
- Who measures. Whether the provider's monitoring or yours counts, and how quickly you have to report an outage.
- Exclusions. Force majeure, faults on the customer side, network problems outside the provider's domain and preview services are almost always excluded.
Service credits are not insurance
If the provider misses the promised level, you are usually entitled to a service credit, a partial refund of the monthly fee for the affected service. For Amazon EC2, for example, it is 10 % when monthly availability falls below the committed level, 30 % below 99 % and 100 % below 95 %, and the claim must be submitted by the end of the second billing cycle after the incident.
Consider an example: the service costs EUR 2,000 per month and an outage lasts 4 hours in a 30-day month. Availability drops to 99.44 % and you receive a 10 % credit, i.e. EUR 200. If an hour of downtime costs you EUR 1,500, the damage is EUR 6,000. A credit is an incentive for the provider, not compensation for you. When choosing a provider, an architecture that prevents outages matters more than the size of the credit.
Availability of the whole system: chains and redundancy
An online shop needs an internet connection, an application server and a database. If any link fails, the shop is down. For components in series, availabilities multiply: A = A₁ × A₂ × … × Aₖ. Three components at 99.9 %, 99.9 % and 99.95 % give 99.75 % together, which means almost 22 hours of downtime per year instead of 8 hours 46 minutes. Every additional component lowers availability.
Redundancy works the other way round. For a group of n identical instances where one is enough, A = 1 − (1 − a)ⁿ. Two servers at 99 % each reach 99.99 % together, and two at 99.9 % reach 99.9999 %. The condition is that they fail independently. Shared power, the same switch, the same faulty software version or a failed automatic failover break this assumption, so the real availability of redundant systems is lower than the calculated one.
A practical tip: find the weakest link first, because improving it lifts the system the most. Often it is not a server but a single internet connection, a single firewall or a database without a replica.
SLA, RPO and RTO
An SLA sets how much downtime is allowed in a period. RTO (Recovery Time Objective) sets how quickly a system must be running again after a single incident, and RPO (Recovery Point Objective) how much data you can lose. The numbers must be consistent: if the RTO for a critical system is 4 hours, one major incident uses more than a monthly 99.9 % SLA allows (43 minutes). A high SLA therefore requires replication and fast failover, not just good backups.
How to set RPO and RTO per system group and check whether your recovery meets them is explained in the guide backup, RPO and RTO.
SLA contract checklist
- Agree with management what an hour of downtime costs and how much downtime per month is still acceptable.
- Map the dependencies of the service and calculate the availability of the whole chain, not just individual components.
- Check the definition of downtime, the measurement window and the measurement method.
- Check the length, frequency and timing of planned maintenance windows.
- Read the list of exclusions and the procedure and deadline for claiming credits.
- Besides availability, ask for support response times and resolution times by priority.
- Align the SLA with RPO, RTO and your disaster recovery plan.
- Set up your own independent availability monitoring and monthly reporting.
Humanfrog runs infrastructure in TIER III data centres in Slovenia and designs redundant architectures and monitoring. If you would like an expert to check the results of the calculator or review an existing SLA, get in touch.