Service level agreement

The commitment

Each site commits to 99.95 % availability in each calendar month, measured against all production servers in that site. Availability is the percentage of the month during which your server was reachable over the network and able to run its workload.

The commitment is per site, not per server, because the failure modes that matter — power, network, a host — are shared by everyone in the site. A single server's own faults (your OS, your software, your configuration) are not availability events.

What counts as downtime

Downtime is any period in which a production server in the site was unreachable due to causes within our control: facility power loss, network failure on our side of the virtual NIC, host failure, or maintenance that was not announced with at least seventy-two hours' notice on the status page.

Excluded from the calculation: announced maintenance, faults caused by your configuration or software, faults caused by third-party services you connect to, denial-of-service attacks beyond our filtering capacity, and force majeure.

The credit formula

For every full minute of downtime in a month, you receive ten minutes of credit, calculated as service time. A ten-minute outage earns 100 minutes — a month of 99.95 % availability earns nothing, and a month of 99.0 % (about seven hours down) earns about three days.

Credits are capped at thirty days per month per server. The cap exists because a site that is down for a month is not a billing problem, it is a refund-and-explain problem, and we would rather do the second.

How credits are applied

Automatically. No claim form, no ticket, no deadline to remember — the credit is added to the affected servers' term end dates within seven days of the month closing, and the status page shows what was credited and to whom.

Credits extend the term of the affected server; they are not cash and cannot be converted to cash or to another asset. The one exception is a deletion request inside the refund window, which is governed by the refund policy.

If a credit has not appeared within fourteen days of month end, write to support with the incident date — we will find it.

Measurement and reporting

Availability is measured from the control plane's own probes plus the hypervisors' heartbeats, sampled every sixty seconds and published per site on the status page.

An incident that pushes a site below 99.95 % for the month is followed by a written post-mortem in the journal within fourteen days: what happened, when, why, what changed, and what was credited. The February 2026 NVMe mirror failure is the template we hold ourselves to.

Support response — not part of the availability number

Tickets are answered by engineers with a median first response of twelve minutes, measured monthly and published. That is a target, not an SLA: missing it costs us embarrassment, not credits, because the remedy for a slow answer is a better queue, not a discount.

Changes

We do not lower the commitment without announcing the change on the status page and by email at least sixty days before it takes effect. Improvements — a higher commitment, a longer measurement window — take effect immediately, because nobody needs notice of good news.