Uptime Figures and SLAs: What the Numbers Promise
Reliability · 10 min read ·
What an uptime percentage means in minutes, what a service level agreement does and does not guarantee, and how to state availability without overclaiming.
Software buyers love a percentage. Ninety-nine point nine feels reassuring; ninety-nine point nine nine feels better. But a percentage with no definition is just a decoration. What is being counted? Over what period? Measured by whom? What happens when it is missed? A published uptime figure is a fact only when those questions have answers on the same page.
This guide explains how availability figures work, what a service level agreement does and does not promise, and how to state uptime in a form that a careful reader can check. The arithmetic is simple, and it is worth knowing, because it makes marketing claims easy to translate into minutes.
The arithmetic
Availability is the proportion of a period during which a service is working. If a service is down for one hour in a thirty-day month, that month had 720 hours and one of them was lost, so availability was roughly 99.86 percent.
Turning percentages into allowed downtime gives a feel for what each "nine" means. For a 30-day month of 43,200 minutes:
- 99 percent allows about 432 minutes, which is over seven hours.
- 99.9 percent allows about 43 minutes.
- 99.95 percent allows about 22 minutes.
- 99.99 percent allows about 4 minutes.
For a 365-day year, 99.9 percent allows about 8 hours and 46 minutes, and 99.99 percent allows about 53 minutes. These are the figures on every availability table, and they come straight from multiplying the period by the unavailable fraction.
The practical lesson is that each extra nine divides the allowed downtime by ten, and the engineering effort required grows faster than that. High availability, as a field, is concerned with designing systems that keep working through failures of individual components. It usually involves redundancy, failover and careful operational discipline, all of which cost money. A vendor claiming four nines for a low price deserves a closer look at the definition.
What is being counted
The headline percentage hides several choices. Ask each of them.
What counts as "up"? Some vendors count a service as available if the login page responds. Others count the proportion of successful requests to the main interface. The first is easy to meet and says little. A good definition names the measurement: for example, the proportion of valid requests to the primary interface that receive a non-error response.
Which components are covered? A product has many parts: the web application, the interface used by integrations, email delivery, background jobs and reports. A figure that covers only the web application may exclude the features a customer depends on most.
What is excluded? Contracts commonly exclude scheduled maintenance, events outside the vendor's control, problems caused by the customer and sometimes failures of third-party platforms the product depends on. These exclusions can be reasonable. They should be listed, and a buyer should check that the list does not swallow the promise.
What is the measurement window? Monthly figures are stricter than annual ones. A vendor can have a poor month and a good year. State the window.
Who measures? Your own monitoring is a start. An independent monitor, or customers' own measurements, is more credible.
SLI, SLO and SLA
Three similar acronyms get muddled, and untangling them improves your writing.
- A service level indicator is a measurement: the share of requests that succeeded, the time to respond to the ninety-fifth percentile of requests, the share of jobs completed on time.
- A service level objective is a target for an indicator: for example, successful requests above a stated share across a month. It is an internal goal and a basis for engineering decisions.
- A service level agreement is a contract. It commits the vendor to a level of service and typically specifies remedies, such as service credits, if the commitment is missed.
An SLA is a subset of what a team measures. The SLO is usually set tighter than the SLA so that the team notices trouble before the contract is breached. When a vendor publishes a target on its fact sheet, say which one it is. "We aim for" and "we guarantee" are different statements and should not share a table cell.
Remedies
Many readers assume an SLA means compensation for any outage. In practice remedies are usually limited. Common features include:
- A credit against future charges, not a cash payment.
- A cap on the credit, often a fraction of the monthly fee.
- A requirement that the customer file a claim within a stated period.
- A sliding scale: more downtime, larger credit.
State the remedy plainly on your page or link to the clause. Buyers who understand the remedy can price the risk. Those who discover it in a dispute feel deceived.
Because this is a contract matter, do not paraphrase beyond what the contract says. Link to the agreement and quote its definitions exactly.
Status pages and incident history
A status page is the public record of how a service performs. A good one has:
- Current status per component.
- A history of incidents with start time, end time, affected components and a short factual summary.
- A subscription option for updates.
- Scheduled maintenance announced in advance.
Link the status page from the fact sheet. If you claim a figure for the past year, the history should let a reader check it. A claim the status page contradicts damages everything else you have said.
After a significant incident, publish a short written review: what happened, how it was detected, how long it lasted, what changed afterwards. Readers rarely expect perfection; they expect honesty and learning.
Related performance claims
Availability is not the only measure of whether a product feels reliable. A service can be "up" and unusably slow. Google's web.dev explains Core Web Vitals, a set of measurements of loading, interactivity and visual stability on the web. If your product is web-based, publishing the response-time target for your interface, alongside the availability target, tells a buyer more than either alone.
Define latency in terms of a percentile rather than an average. "Ninety-five percent of requests complete within a stated time" survives scrutiny; "average response is fast" does not, because averages hide the slow tail that real users feel.
Writing the availability row of a fact sheet
A useful row has these elements:
- Target or commitment: state which it is.
- Measure: the indicator in one sentence.
- Window: monthly or annual.
- Components covered.
- Exclusions: link to the clause.
- Remedy: link to the clause.
- History: link to the status page.
- Last reviewed: the date.
An example, in prose: "Contractual availability target for the web application and public interface, measured as the proportion of successful requests per calendar month, excluding announced maintenance and customer-caused errors. Credits apply as described in the agreement. History is published on the status page." That is dull, and complete.
Mistakes to avoid
Rounding up. If last year was 99.87 percent, do not say 99.9. Say what it was.
Quoting the infrastructure provider's figure. The provider's availability is not yours. A product built on a platform with a strong target can still fail through its own software or configuration.
Counting only whole-service outages. Partial degradations still hurt users. If you only count total outages, say so.
Forgetting maintenance. If maintenance windows regularly take the product offline, a buyer in a different time zone will experience them as downtime, whatever the contract says.
Worked example: reading a vendor claim
Imagine a vendor page that says: "99.95 percent uptime guaranteed." Here is how a careful reader turns that sentence into questions.
First, the arithmetic: over a 30-day month, 99.95 percent allows about 22 minutes of unavailability. Is that credible for a product with this vendor's size and history? A look at the status page for the last twelve months shows whether incidents of that length are rare or routine.
Second, the word guaranteed. Guaranteed by what? If the agreement says credits are available only for months in which availability falls below a lower threshold, then the guarantee is weaker than the headline. If credits are capped at a small share of the monthly fee, then a bad month costs the vendor little and the buyer a great deal.
Third, the scope. Does the figure cover the interface that the customer's integration will use, or only the web application? Do background jobs count? Does the percentage apply to every plan or only to the highest?
Fourth, the measurement. Who records the downtime? If the vendor's own monitor decides, ask how it is set up and whether it checks from outside the vendor's network.
None of these questions is hostile. A confident vendor can answer them in a paragraph, and the answers become part of the buyer's decision record.
What to do if you cannot meet the target you want to claim
Smaller teams sometimes feel pressure to match the figures of larger competitors. Resist it. A modest, accurate target is a stronger position than an ambitious one you cannot meet. Publish what you achieved last year, describe how you plan to improve and say plainly that you offer no contractual commitment if that is the case. Some buyers need a contract and will move on; many others simply need to know the truth, and will respect a team that tells it.
If you do want to improve, the usual path starts with measurement: define the indicator, record it for a quarter, find the largest causes of downtime and fix those in order. Redundancy and automated recovery come after you know where the failures actually are. Improving what you have measured is far cheaper than buying nines you have not earned.
Summary
Availability numbers translate into minutes, and every extra nine divides the allowance by ten. A figure is a fact only when it names the measure, the window, the covered components, the exclusions and the remedy. Keep SLI, SLO and SLA distinct, link to the status page, and report the real figure rather than a rounded one. Readers will forgive an honest 99.8 far more readily than a 99.99 they cannot reconcile with the status history.
Questions and answers
- How much downtime does 99.9 percent allow?
- Over a 30-day month, 99.9 percent availability allows about 43 minutes of downtime. Over a full year it allows about 8 hours and 46 minutes.
- What is the difference between an SLA, an SLO and an SLI?
- An SLI is a measured indicator, such as the proportion of successful requests. An SLO is an internal target for that indicator. An SLA is a contractual commitment, usually with remedies if it is missed.
- Does a 99.99 percent claim mean the product never goes down?
- No. It means that over the measurement period, downtime stays within a very small budget, as defined by the contract, and subject to its exclusions.
- Where should a vendor publish uptime history?
- On a public status page that records incidents with start and end times, affected components and a short explanation, linked from the fact sheet.