All error codes

no_gpu_capacity

HTTP 409 Nothing free right now fits this request

What happened

The platform has nothing free that fits the request, or the one host you chose has nothing free. WHEN YOU CHOSE THE HOST YOURSELF (the host field), this code means that host alone: the detail field names it and says what it lacks, and the server is NEVER moved to another host silently - that was the point of choosing. A host that takes no new servers at all answers the same code. SEVERAL CAUSES SHARE THIS CODE on a type where you pick the cards, and the detail field of this response says which one it was, in words. Four are shortages: the cards themselves, the memory reserved on the host that has them, the local disk on that host, and hosts with no local storage at all when you asked for a local disk. Three are passing states of the platform: the host that could take the cards is not answering, the cluster behind the type is not available, or no host of the type is ready to take a server. The detail is worked out from the platform inventory as it reads at the moment of the answer, so a concurrent order may have changed it since the refusal; when it cannot be worked out, the field is absent rather than guessed. Resizing a GPU Pro server answers this code too, when the host it runs on has no reserved memory left for the new size; the detail field then says so. On a type sold as a GPU package there is only one shortage to have: no whole host is free - a package is one host with its card, and hosts are counted in single digits. Your request is correct; cards and hosts are not shared between servers, so a shortage lasts until someone releases theirs.

What to do

If you chose the host yourself, the detail names that host: order fewer cards or less memory or a smaller disk, name another host from GET /v1/server-types/{id}/hosts, or drop the host field and let the platform place the server. On a type where you pick the cards, read the detail first and change what it names: fewer cards or another model when it says the cards, less memory when it says the memory reserve, a smaller disk or disk_type ceph when it says the local disk, disk_type ceph when it says there is no local storage - the type's pricing lists the models and the limits. When the detail says the host is not answering, repeat the same request in a moment; when it says the cluster is not available or no host is ready, repeat it later. A refused resize leaves the server as it was on its host: ask for less memory. On a package type there is nothing to adjust in the request: wait for a host to free up, or create the server on another type. For a shortage, repeating the same request unchanged frees nothing, so retry only when the request changes or you know capacity was released. A refused request costs you no address: one it asked for as new is released again, and an address you named yourself (or one reuse found free) was already yours and stays, billed as before.

Retrying

Change the state this refers to, then repeat the same request with the same idempotency key.

How it arrives

{
    "type": "https://docs.fiberax.com/errors/no_gpu_capacity",
    "title": "Nothing free right now fits this request",
    "status": 409,
    "code": "no_gpu_capacity",
    "request_id": "a277dcd1ca652c65"
}

Branch your code on code, never on the wording of title.