One page, four minutes to read, and no clause that makes the commitment evaporate on the month you need it. This agreement applies automatically to every instance on the network, at every plan, in every city, with no tier that has to be purchased first.
Contractual monthly uptime is 99.99 percent per instance. Measured uptime across the last twelve rolling months is 99.993 percent, published continuously on the status page (/status).
1. What counts as downtime
Downtime is the period during which your instance is unreachable from at least two of our external probes, both located outside the affected site, where the cause lies on our side of the demarcation.
Probes sample every thirty seconds. A downtime event opens at the first of two consecutive failed samples and closes at the first successful one, which means the shortest event we can record is sixty seconds. Anything shorter than that is not counted, in your favour or ours.
Loss of a service inside your instance is not downtime. A machine that answers on the network while your database is wedged is up, however unhelpful that distinction feels at the time.
2. What is excluded
- Scheduled maintenance. Announced at least seventy-two hours ahead on the status page and by email, capped at four hours per instance per calendar month, and scheduled between 02:00 and 06:00 local site time. Maintenance beyond that cap counts as downtime like anything else.
- Emergency maintenance. A kernel or hypervisor fix that cannot wait, announced as far ahead as the fix allows. Roughly two of these happen a year.
- Customer-caused outages. Your kernel panic, your firewall rule, your out-of-memory killer, your expired renewal, or a suspension under the acceptable use policy (/legal/aup).
- Attacks beyond the filtering tier. Volumetric traffic aimed at your service above the scrubbing capacity of the site, which is up to 12 Tbit/s at the largest sites and between 4 and 8 Tbit/s elsewhere. Layer-7 attacks are excluded unless you hold DDoS Pro, which exists precisely for that case.
- Things past our edge. A failure of one transit provider is ours, because we buy several and route around it. A break in the public internet between your users and our edge is not.
- Force majeure, in the ordinary sense of the phrase.
Everything not on that list counts. Hardware faults count, hypervisor faults count, our own configuration errors count, and a site that loses power counts even though the power was somebody else’s job.
3. Credits
| Measured monthly uptime | Downtime in a 30-day month | Credit |
|---|---|---|
| 99.99% or above | up to 4 minutes | none due |
| 99.90% to 99.98% | 4 to 43 minutes | 10% of the monthly fee |
| 99.50% to 99.89% | 43 minutes to 3 hours 36 minutes | 25% |
| 99.00% to 99.49% | 3 hours 36 minutes to 7 hours 12 minutes | 50% |
| below 99.00% | more than 7 hours 12 minutes | 100% |
Credit is calculated per instance against that instance’s monthly fee, at the effective monthly rate of your billing cycle rather than the list price. Add-ons attached to the affected instance are included in the fee the percentage is applied to. The maximum in any month is one hundred percent of that fee.
4. How credits are applied
Automatically, and without you noticing beforehand. When a calendar month closes, the billing job compares every instance’s measured uptime against the table above, posts any credit to the account balance in euro within three days, and prints a line on the next invoice explaining which instance and which month.
There is no claim form, no thirty-day deadline to spot the outage yourself, and no requirement to have opened a ticket during it. Roughly one credit in five that we issue goes to a customer who had not realised anything happened.
Credits offset future invoices. They are not paid out to a chain address, for the reason set out in the refund policy (/legal/refunds): no money entered the company against them.
Service credits are the sole remedy for downtime, as section 9 of the terms (/legal/terms) sets out.
5. Network targets
Uptime is the headline, though it is rarely the thing that ruins a week. Three further targets apply to our own network, meaning from your instance to the edge of our backbone (/network):
- Packet loss below 0.1 percent, sampled every thirty seconds and averaged across a calendar month.
- Jitter below 2 milliseconds between sites within the same region.
- Latency within 20 percent of the site-to-site medians published on the location index (/locations), which come from our own probes rather than a brochure.
Missing any of these for a full calendar month on a path your instance uses is treated as a 99.90 to 99.98 band event and credited at ten percent, on the same automatic basis.
6. Measurement and disputes
Our probes and the looking glass (/network/looking-glass) are the primary record, and the raw samples for your instance are available on request for the last ninety days.
Disagree with a month? Send your own monitoring data. If it samples at least once a minute from at least two independent networks, we will take it seriously, and we have credited on customer figures more than once. Nobody at this company has ever argued a customer out of a credit worth less than the argument.
This agreement forms part of the terms of service (/legal/terms) and changes on the same notice period.