Almost every server, hosting and cloud service offer includes a number: 99.9%, 99.95%, 99.99%. It comes from the provider’s SLA. SLA stands for Service Level Agreement: the part of the contract that records, in a measurable way, what the provider commits to and what happens if it fails to deliver. In this article we show what the numbers mean in minutes, what parts an SLA has, and what to look for when you read the SLA of a cloud provider or an operations partner.
What an SLA means in IT operations: what does it contain?
A good SLA answers at least these questions:
- What does it cover? The server, the network, the application, or only the hardware?
- What availability is committed to, and over what period is it measured (monthly, yearly)?
- What counts as “unavailable”, and who measures it, and from where?
- Response and resolution times per incident priority, and in which period (business hours or 24/7).
- Maintenance windows and how you are notified.
- Exclusions: what doesn’t count as downtime.
- Service credit: what you get if the commitment is not met, and how to claim it.
Availability: how many minutes of downtime does a percentage allow?
Percentages on their own say little, so it’s worth converting them. The table assumes a 30-day month and a 365-day year:
| Availability | Allowed downtime per month | Allowed downtime per year |
|---|---|---|
| 99% | 7 h 12 min | 3 days 15 h 36 min |
| 99.5% | 3 h 36 min | 1 day 19 h 48 min |
| 99.9% | 43 min 12 s | 8 h 45 min 36 s |
| 99.95% | 21 min 36 s | 4 h 22 min 48 s |
| 99.99% | 4 min 19 s | 52 min 34 s |
| 99.999% | 26 s | 5 min 15 s |
Two things are worth noticing. One is the measurement period: 99.9% measured over a year allows a single outage of almost nine hours, but 99.9% measured monthly does not. The other is the chain: if your service depends on two components that build on each other (for example a server and a database), and both are at 99.9%, the combined availability is only around 99.8% (0.999 × 0.999). Every additional dependency pulls it lower.
Response time and resolution time
The two are often confused, but they are very different:
- Response time: how long it takes for a specialist to start working on your incident. An automatic “we have received your ticket” email is not a response.
- Resolution time: how long it takes for the service to be restored. Fewer providers commit to this, and when they do, it is often only as a “target”.
Both differ by priority. Here is an example of what this usually looks like:
| Priority | Example | Response time | Period |
|---|---|---|---|
| Critical | The production system is unavailable | 30 minutes | 24/7, weekends included |
| High | Slowdown, a feature doesn’t work | 2 hours | Business hours |
| Normal | Request, minor bug | 1 business day | Business hours |
Watch the period: a four-hour response time in business hours means Monday morning for an incident reported at 16:00 on a Friday.
Maintenance windows and exclusions
Planned maintenance is almost always left out of the availability calculation. That’s fine, but the contract should state when maintenance can take place (for example between 22:00 and 06:00 on weekdays), how far in advance you are notified, and how much of it is allowed per month. Otherwise, even with a 99.9% SLA, there can be a four-hour “planned” outage every week.
Typical exclusions are force majeure, the client’s own error or intervention, internet problems outside the provider’s network, and suspension of the service (for example for non-payment). These are reasonable, but the more broadly they are worded, the less the committed number is worth.
Service credit: what do you get if the commitment is not met?
A breach of the SLA almost never means compensation for damages. The provider credits part of the monthly fee for the service in question, usually against future invoices, and only if you ask for it, within the deadline, with evidence. It doesn’t cover your lost revenue or lost customers – we wrote about this in more detail in our article on the impact of service outages.
How to read a cloud provider’s SLA: the AWS EC2 example
The Amazon EC2 SLA (the version of 25 May 2022) is a good illustration of what to look for. It contains two commitments:
| Region-Level SLA | Instance-Level SLA | |
|---|---|---|
| Condition | The running instances are in at least two Availability Zones (AZs) at the same time | A single EC2 instance |
| Committed monthly availability | 99.99% | 99.5% |
| 10% credit | Between 99.0% and 99.99% | Between 99.0% and 99.5% |
| 30% credit | Between 95.0% and 99.0% | Between 95.0% and 99.0% |
| 100% credit | Below 95.0% | Below 95.0% |
The lessons:
- A single virtual machine is only covered by 99.5%, which allows 3 hours 36 minutes of downtime per month. You only get 99.99% if your machines run in several zones, and a region-level outage only counts if all your instances in all zones are unavailable at the same time.
- “Unavailable” = no external connectivity. If the machine is reachable but your application doesn’t work, that doesn’t count.
- The credit is a percentage of the EC2 bill, and it can only be applied to future EC2 charges. In addition, AWS doesn’t charge for any hour in which an instance was unavailable for more than six minutes.
- You have to claim it: by the end of the second billing cycle, with a case opened in the AWS Support Center, with request logs that prove the outage.
It’s worth following the same logic with every provider: what is the condition, what is measured, how much is the credit, and what do you have to do to get it.
SLA, SLO, SLI – what’s the difference?
- SLI (Service Level Indicator): the measured metric, for example the proportion of successful HTTP requests or the 95th percentile of response time.
- SLO (Service Level Objective): the internal target for this metric, for example 99.95% successful requests per month.
- SLA: the contractual commitment, with consequences. It is usually looser than the SLO, so that there is a margin before credits have to be paid.
If you run your own service, it’s worth setting SLOs even if you don’t commit to an SLA with anyone: they show where improvement is needed. We wrote about a few proven methods in our article on infrastructure reliability.
What to ask an operations partner
- What does the availability cover: the hardware, the operating system, or the application too?
- How and from where is it measured, and can you see the results?
- What are the response and resolution times per priority, and is there a 24/7 on-call service?
- How can you report an incident (phone, ticketing system), and who picks up at night?
- When can maintenance take place, and how far in advance are you told?
- How quickly, and to what state, is the system restored after data loss (RTO, RPO)?
- Do you get a monthly report on outages and tickets?
An SLA is worth as much as the practice behind it: monitoring, on-call, tested restores. We agree with our clients on what we commit to, and within what time, based on how the system is actually built. You can read about our operations services on the Infrastructure and operations page.