Deutsche Cloud
Menu

Service level agreement

What we commit to for the console and the API, how we measure it from outside, what you get back when we miss it, and the drill numbers it rests on.

Draft: not in effect yet. It applies to usage from a date published here.

Commitment

Availability
99.9 % per calendar month
Allowed downtime
about 43 minutes in a 30-day month
Covered
The console, the public API and the CLI, Terraform and MCP access through it, and the Git, npm and container-registry endpoints the platform serves
Not covered here
Your machines, databases, apps, buckets and other resources. They keep running when the console is down; their own service levels are a separate page, not yet published.

How downtime is measured

  • An uptime probe in Helsinki, outside the platform it watches, requests the console’s health endpoint every 30 seconds with a 10-second timeout.
  • A minute counts as unavailable when the probe gets no successful answer in it. When a second probe location runs, a minute counts only if both see the failure; until then the one probe is binding.
  • Monthly availability is the share of available minutes in the month. It is the probe’s number, not our server logs.
  • Your own evidence (times, request IDs) is accepted in a claim. It can add unavailable minutes, never remove them.

Service credits

If a month’s availability falls below 99.9 %, your organization receives a credit on its next invoice, as a share of that month’s whole invoice:

Monthly availability below Credit
99.9 %10 %
99.0 %25 %
95.0 %50 %
  • Credits are capped at 50 % of the month’s invoice and are set against the next invoice. They are not paid out.
  • Claim within 30 days after the end of the month by writing to info@deutschlandgpt.de, naming your organization and the month.
  • Free and trial accounts are measured like everyone else and see the same numbers, but receive no credits.

Exclusions

Minutes in these situations do not count as unavailable:

  • Scheduled maintenance announced on the status page at least 5 business days ahead, outside 08:00–20:00 CET, at most 2 hours a month. Deploys and planned database switchovers are not maintenance: they are counted.
  • Unavailability caused by you: your credentials, your configuration, abuse from your account.
  • Force majeure in the narrow sense: war, natural disaster, action by a state authority.
  • Features marked as preview.

Counted, not excluded

An outage at our hosting provider, including the loss of a whole data-centre location, and failures of upstream networks or DNS providers count as downtime. Today two of the platform database’s three machines are in Falkenstein, so losing that location stops the console, and that time counts as downtime too.

Recovery targets

For the platform’s own database. These are targets, not credited commitments: they say what our design and drills support.

Failure Data loss (RPO) Back in service (RTO)
One database machine lostnone≤ 1 minute
The whole database cluster lost≤ 5 minutes≤ 1 hour

What the drills measured

13 November 2026

On a copy of the production layout (a load balancer, two control-plane machines in Falkenstein and Nuremberg, a three-machine PostgreSQL cluster), measured from outside with a request every 250 milliseconds and a write every 500 milliseconds:

Failure What clients saw Writes lost
A control-plane machine powered offno failed request; some answers took about 5 s for 8 s0
The app on one machine stoppedabout half the requests failed for 9 s0
The database leader’s machine powered offevery request failed for 30 s (44 s with the database’s default failover timing)0
Planned database switchoverevery request failed for 3 s0
All three database machines lost, restored from backup2.5 minutes, with the steps preparedthe last 4.6 minutes
The Falkenstein location lostdown for as long as it was, plus 40 s0

The service level agreement in the documentation

Measured availability

Published here for every month from the date this SLA takes effect.