Australian cloud for business software — built and run in Sydney since 2002

NetLeverage
NetLeverage

Reading an availability figure in a service report: what was measured, from where, and what was left out

Technical article
Hosting

About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.

An availability figure such as 99.98 per cent is the output of a definition, not a measurement of your experience. Three choices sit behind it: what counts as "available", where and how that is measured, and which outages are removed before the percentage is calculated. Change any one and the same month produces a different number.

So the useful question about a report is not 'is 99.98 good?' but 'what was measured?' The rest of this guide shows how published cloud agreements answer that, so you can recognise the same pattern in a report from any provider. It does not repeat the arithmetic of chaining components, which is covered in an existing article on availability arithmetic.

Why it matters

Take the figure at face value and 99.98 per cent over a 30-day month is about 8.6 minutes of downtime (our arithmetic: 43,200 minutes multiplied by 0.0002). A business that suffered a 40-minute period when staff could not sign in will find that hard to square with the report.

Microsoft's guidance on reading agreements warns against the shortcut: "A 99.9% SLA doesn't guarantee that the service always remains available" and the document says the figure's definition of downtime "determines what gets measured" (Microsoft, How to read a service-level agreement). Narrow definitions mean problems that affect users may not breach the figure at all.

A report is also usually a monthly summary. Microsoft notes: "SLAs usually measure uptime over a billing period, not in real time." A reported month can include a visible outage and still look strong.

What counts as available

Different providers define downtime differently, and the differences are in plain text in their agreements. The following examples are published agreements for infrastructure services. They are used here only to show how definitions vary, and they say nothing about any particular hosting provider.

  • Amazon EC2, instance level: unavailability means "your Single EC2 Instance has no external connectivity" (AWS, Amazon Compute SLA, last updated 25 May 2022). For the regional commitment, it means that all running instances in two or more zones "concurrently have no external connectivity".
  • Google Compute Engine: downtime for a virtual machine means "loss of external connectivity or persistent disk access for the Single Instance" (Google Cloud, Compute Engine SLA, last modified 14 October 2026). So loss of disk access counts, which is broader than network reachability alone.
  • Microsoft Azure Virtual Machines: the measure is "Virtual Machine Connectivity", defined as "bi-directional network traffic between the Virtual Machine and other IP addresses using TCP or UDP network protocols in which the Virtual Machine is configured for allowed traffic" (Microsoft, Service Level Agreement for Microsoft Online Services, document dated 1 October 2026).

What these share is that availability means the machine can be reached, not that your application works. A server can be reachable while its database is stuck, its login service is down or its disk is full. In none of the three definitions above would that necessarily count as downtime. That is our reading of the wording, not a statement about any service's behaviour.

A report that says 'availability of the hosting platform' may therefore be describing the layer beneath your application. Ask which layer.

Time-based or request-based

Microsoft's guide distinguishes two methods: "Time-based measurement typically tracks whether the service was available during each time interval. Request-based measurement tracks the ratio of successful responses to total valid requests." It adds a warning that matters for any report: "The same outage can produce different availability percentages depending on which method the SLA uses."

Amazon S3's agreement is an example of request-based measurement. Errors are counted for each 5-minute interval, and the monthly figure averages them; the agreement adds that "If you did not make any requests in a given 5-minute interval, that interval is assumed to have a 0% Error Rate" (AWS, Amazon S3 SLA, last updated 28 November 2023). So an outage in the middle of the night, when nobody makes requests, would not move that figure (our reading).

Time-based measurement has its own edges. Google's agreement says a "Downtime Period" is one or more consecutive minutes of downtime, and "Partial minutes or intermittent Downtime for a period of less than one minute will not count towards any Downtime Periods". A service that fails for 30 seconds every few minutes might record no downtime at all.

When you read a report, find out whether the figure counts minutes, requests or users, and the smallest unit that counts.

Where it was measured

The same service looks different from different places. Google's site reliability guidance puts it this way: many measurements are gathered on the server side, but "not measuring behavior at the client can miss a range of problems that affect users but don't affect server-side metrics" (Google, Site Reliability Engineering, chapter 4). The same chapter says that "raw data is collected over a measurement window and then turned into a rate, average, or percentile", which is another way of saying that the window chosen shapes the result.

For a hosting report, the useful questions are:

  • Is the figure produced by the provider's monitoring, from inside its own network, or from a probe outside it?
  • Does the probe test a ping, a port, or a real page or login?
  • How often does it test? A check every five minutes cannot see a three-minute outage reliably (our reasoning).
  • Does the same tool count planned work?

A figure measured inside the provider's network describes the provider's platform. A figure from an external check of your own application is much closer to what your users experience.

What is excluded before the percentage is calculated

Agreements remove some outages from the calculation. Microsoft's guide lists the common categories: SLAs "commonly exclude downtime from planned maintenance, force majeure, customer-initiated actions, misconfiguration, exceeded quotas, malformed requests, and preview or unsupported features".

The published agreements show this in practice:

  • The AWS agreements exclude issues "caused by factors outside of our reasonable control", including internet problems beyond the service's demarcation point, and issues resulting from the customer's own actions, equipment or software.
  • The Google agreement excludes pre-general-availability features and errors "caused by factors outside of Google's reasonable control" or resulting from the customer's software or hardware.
  • The Microsoft agreement (October 2026 edition) states that "Monthly maintenance window that incurs a downtime to patch your server and infrastructure is excluded from the uptime calculation", and that customer-initiated operations such as restart or scale "that incur downtime are excluded from the uptime calculation". It also defines Scheduled Downtime as downtime for maintenance or upgrades, with the promise "We will publish notice or notify you at least five (5) days prior to the commencement of such Downtime."

So a month with a planned maintenance outage can report 100 per cent. That is not dishonest if the exclusion is stated. It is misleading if you read the figure as 'the service was up'.

Ask for two numbers: the availability figure as calculated under the agreement, and the total minutes of any interruption you could have noticed, including planned work.

Aggregation: one server or many

When a figure covers several servers, check how they are combined, because the method decides what the number can hide. Microsoft's guide tells readers to determine "whether the SLA measures availability per individual service instance or across all instances in a scope", and warns that "healthy instances can offset failed instances".

Two different methods are easy to confuse:

  • An average or pooled figure. A report for 'your environment' might average four servers, one of which was down for much of the month. An average of this kind can never be higher than the best server, but it can hide the affected server behind three healthy ones. If the server that was down is the one your staff use, a single combined figure of 99.9 per cent says little about their month. With equal weighting, one server down for the whole of a 30-day month among four servers still produces a combined figure of 75 per cent, and one down for a day produces about 99.2 per cent, which can look acceptable on a monthly report.
  • An any-instance-available metric. Some services define the service as available if at least one instance of a redundant group can answer. That is a different metric from an average. Because a healthy instance covers for a failed one, it can legitimately be higher than the availability of any single instance. It measures whether the service as a whole stayed reachable, not whether each server did.

Ask which of these a report uses. (The Microsoft guide also says the offsetting "might produce a higher reported availability than any single instance actually experienced". That holds for an any-instance-available metric. It does not hold for an ordinary average, so we have not relied on the sentence for any arithmetic here.)

Ask whether figures are given per service and per server, or only as one number.

What to choose and what to agree

This section is our recommendation. For each availability figure in a report, agree these in writing:

  1. The definition of available. For example, reachable on the network, responding to a specific page, or able to complete a sign-in.
  2. The measurement point. Inside the provider's network, or an outside probe, and how often.
  3. The smallest unit that counts. For example, one minute, five minutes or one request.
  4. The exclusions. A list, with planned work shown separately and not silently removed.
  5. The period and the aggregation. Monthly or other, per server or combined.
  6. Your own measurement. An independent check of one user journey, so you hold a second figure.

Read-only checks

These checks read documents and reports and change nothing. No commands are included in this article.

  • Find the paragraph in your agreement that defines downtime or unavailability and compare it with the report's wording.
  • Find the list of exclusions and ask whether planned maintenance is included in the report as a separate item.
  • Compare the report's outage list with your own helpdesk records for the month.
  • Check whether the report is produced by an automated system or compiled by hand.

A worked example (illustrative)

The following is invented. A business receives a monthly report showing 100 per cent availability for its hosted application server. During the month its staff could not sign in to the accounting application for 40 minutes one morning. When the business asks, the provider explains that the report measures whether the server answers a network check every minute, and the server answered every check, so on that definition there was no counted downtime and the figure is 100 per cent. The accounting application's login service had stopped, which the network check could not see.

Nothing in the report was false. The definition of available was network reachability. The business asks for a second figure, from a scheduled check that opens the login page and signs in with a test account, and agrees to receive both numbers each month.

Risks and limits

  • The agreements quoted above belong to large infrastructure providers and are examples of how definitions vary, not descriptions of any hosting provider's reports.
  • Agreements change. The Microsoft document is a consolidated agreement, the October 2026 edition, and we read only its Virtual Machines and general sections.
  • We have not seen the reports of any specific provider; the questions here apply to any.
  • This article explains reading a figure. It does not argue that any specific figure is too low or too high.

Next step

Ask your provider for the written definition behind the availability figure in your latest report, including the measurement point, the smallest counted unit and the exclusions. Then compare the report's outage list with your own records for the same month. If you are reviewing a hosted service, the Managed Services page is a starting point for that conversation; ask for the scope in writing.

Sources

All accessed 15 November 2026.

Learning notes

  • An availability percentage is the result of a definition; read the definition first.
  • Reachable is not the same as working.
  • Exclusions, especially planned maintenance, are removed before the percentage is calculated.
  • Averaging across servers or intervals can hide a bad one.
  • A second figure from your own check is worth having.