It turns a service-level agreement into numbers you can act on. You see the downtime the percentage allows per day, month and year, the ceiling your dependency stack lets you promise, and what a breach costs in credits. This page does all three, and drafts the clause itself, since the wording decides as much as the number. For the definitions behind the terms, see our “What is SLA & uptime?” guide.

Start from the ceiling, not the ambition. Multiply the SLAs of everything you depend on (the ceiling section of the result does it live), subtract a margin for your own failures, and promise the result. Say the chain caps you at 99.82%. A 99.5% you always beat builds more trust than a 99.9% you breach twice a year. Breaches cost credits, renewal conversations, and the one thing an SLA is really for: credibility.

If your service needs A and B to work, your best case is both being up at once: multiply them. Hosting 99.95% × DNS 99.99% × payments 99.9% × CDN 99.98% = 99.82% — that’s your ceiling before your own code fails at all. The multiplication is unforgiving: every serial dependency, however solid, lowers it. The way back up is parallel redundancy: two independent 99.5% instances give 1 − 0.005² = 99.9975%. That holds only while they fail independently — same region or same deploy pipeline means one failure domain, not two.

Tiered by how badly the month missed. Commonly 10% of the monthly fee for a miss below the target, 25% for landing more than a tier below, and 50% (often the cap) for a catastrophic month. Two constants run through nearly all of them. Credits are the sole remedy — not refunds, not damages. They must be claimed within a deadline, usually 30 days, and with evidence. The exposure section of the result prices what that schedule costs you as the provider.

Neither, if you can negotiate. A yearly window gives 99.9% an 8h 45m pool for the whole year, so an 8-hour outage in March still counts as met when the rest of the year is clean. Under a monthly window, that outage is 8 hours against a 43m 48s budget — a breach in its own month. Maintenance exclusions are reasonable when capped and announced. Check how many hours per month, and with how much notice. Uncapped “planned maintenance” turns any percentage into decoration.

Whichever you pick, name it in the clause — “99.9%” without a clock is an argument waiting to happen. Business-hours SLAs cover 45 of the 168 weekly hours (Mon–Fri 9–18). That makes the same percentage stricter by day and blind at night: a weekend outage scores 100%. Our uptime calculator walks one outage through both denominators — the grade moves from “met” to “breached” without a single number in the SLA changing.

No — and they’re not meant to. The asymmetry runs both directions. As a customer, your €49 host owes you €4.90 for an outage that cost your shop €14,000. As a provider, the credits you owe are the smallest line of your breach month — the real bill is churn and the renewal discount you’ll concede. Credits are a signal of confidence and a pricing mechanism. Redundancy and fast detection are the actual insurance.

With measurements neither side controls. The provider’s own dashboard is the other party’s evidence. Your claim, or your defense, needs independent timestamped checks from outside the infrastructure in question. That’s also the cheapest way to end the argument before it starts: publish the numbers on a public status page, and both sides read the same ledger. Uptimia does both — monitoring from 171+ locations in 70+ countries and a public status page fed by it.