Scheduled Maintenance and the SLA Math Nobody Checks

Read the definitions section of any vendor SLA and you will find that availability is measured over a period from which scheduled maintenance has already been removed. The commitment is not 99.9% of the month. It is 99.9% of the month minus whatever the vendor announced in advance.

This is not a trick and it is not unreasonable. Systems need maintenance, and a vendor that could never take a window would ship more slowly and less safely. What makes it worth attention is that the exclusion is almost never quantified in the same document. The SLA states a percentage and defines maintenance as excluded, and then says nothing about how much maintenance is permitted, which leaves the effective ceiling undefined in the contract you just signed.

The arithmetic below shows what different allowances do to a headline number, why emergency maintenance is the clause that actually costs you, and what to ask for during negotiation. It is for anyone who has to publish an availability target downstream of vendors, or who has ever been surprised that a quarter with no SLA breach contained several hours of unavailability.

Take a 30-day month, which is 720 hours, and a contractual 99.9%. With no maintenance exclusion the vendor may be unavailable for 43 minutes. Now exclude four hours of announced maintenance. The 99.9% applies to the remaining 716 hours, permitting about 43 minutes of unplanned failure on top of the four hours you already agreed to lose.

Your actual experience is 4.7 hours of unavailability across the month, which as a fraction of the real 720 hours is 99.34%. The contract is fully honoured. The number you can rely on is 99.34%, and the downtime behind it is more than six times what the headline implies. If you are composing several vendors, the adjusted figure is the one that belongs in the calculation described in dependency chain availability math.

Scheduled maintenance is at least bounded and announced. The clause worth reading twice is the one covering emergency or unscheduled maintenance, which many agreements also exclude and which by its nature carries no notice period and frequently no cap.

An exclusion for emergency maintenance with no ceiling on frequency or duration makes the availability commitment unbounded in the vendor's favour, because any unplanned outage can be characterised as emergency maintenance after the fact. Most vendors do not abuse this and a few do, and the difference is not visible until you try to claim a credit and find the incident reclassified.

Ask two questions during negotiation. Is emergency maintenance capped, by frequency or by total hours per period? And who determines the classification, with what notice? A vendor unwilling to cap it at all is telling you something useful about how they think about the commitment.

Watch for the related pattern where maintenance is announced with technically compliant but practically useless notice. A window announced seven days ahead is plannable. The same window announced four hours ahead is an outage with paperwork, and it satisfies a clause that only requires advance notice without specifying how much.

Negotiate the cap before the percentage. Moving a vendor from 99.9% to 99.95% is a harder conversation and worth less than capping maintenance at two hours a month, which costs the vendor little if their real usage is below the cap and protects you if it is not.

Ask for windows outside your peak. A maintenance window is far cheaper during your quietest hours, and vendors serving multiple regions can often accommodate a preference when asked at contract time. Nobody offers this unprompted.

Require notice in the contract with a number attached. Advance notice as an undefined term is worth very little; seven days is a reasonable ask for planned work and gives your team time to shift its own deploys away from the window.

Then measure what actually happens, because none of the above tells you what the vendor does. Track announced windows, whether they started and ended when announced, and whether availability during the window matched what was announced. Windows that overrun are the most common quiet breach, and they are invisible to anyone reading only the SLA report the vendor produces. When something is genuinely claimable, the SLA credits calculation guide covers the process.

If your vendors exclude maintenance from their numbers and you exclude maintenance from yours, the exclusions compound and your published figure drifts further from what customers experienced. Customers do not experience exclusions. They experience whether the thing worked when they tried it.

Publish both figures. An SLA-basis number governs credits and belongs in the contract conversation, and a raw wall-clock number describes the service as lived. Showing both is more credible than showing the flattering one, and it removes the awkward conversation where a customer's own measurement disagrees with your report and neither side can reconcile the gap.

Keep the maintenance log alongside the availability history. A year of announced windows with actual start and end times answers questions that a percentage cannot: whether maintenance is trending up, whether it lands in your peak, and whether a particular vendor's windows reliably overrun. The definitional groundwork is in understanding SLA metrics.

Does scheduled maintenance count against an SLA? In almost all commercial SLAs, no. Availability is measured over a period with announced maintenance removed, so time inside an announced window is not counted as downtime and does not contribute to a credit claim.

How much does a maintenance exclusion actually cost? More than most people estimate. Four hours a month against a 99.9% commitment lowers the effective ceiling to 99.34%, which is 4.7 hours of real unavailability instead of 43 minutes. Eight hours a month brings it below 98.8%.

Can a vendor call an unplanned outage emergency maintenance? If the agreement excludes emergency maintenance without defining or capping it, then in practice yes, and you may find out when a credit claim is refused. Cap it at contract time, and define who classifies an event and on what notice.

Should I exclude maintenance from the availability I report to customers? Report both. Use the SLA basis for contractual questions and a raw wall-clock figure for what customers actually experienced. Publishing only the flattering number invites a dispute you cannot win, because the customer's own measurement is the one they trust.

What maintenance cap is reasonable to ask for? Two hours a month with seven days notice is achievable from most mature providers and costs them little if their real usage is already below it. A vendor who will not cap it at any number is worth asking why.