[CI][C++] CUDA docker images fail building

[asfimport]

We have Ubuntu-based CUDA-enabled builds using Docker images that evidently have not been exercised in a long time:

$ archery docker run ubuntu-cuda-cpp
Pulling ubuntu-cuda-cpp ... done
WARNING: Some service image(s) must be built from source by running:
    docker-compose build ubuntu-cuda-cpp
Building ubuntu-cuda-cpp
[+] Building 0.9s (4/4) FINISHED                                                                                                          
 => [internal] load build definition from ubuntu-20.04-cpp.dockerfile                                                                0.0s
 => => transferring dockerfile: 5.39kB                                                                                               0.0s
 => [internal] load .dockerignore                                                                                                    0.0s
 => => transferring context: 35B                                                                                                     0.0s
 => ERROR [internal] load metadata for docker.io/nvidia/cuda:9.1-devel-ubuntu20.04                                                   0.9s
 => [auth] nvidia/cuda:pull token for registry-1.docker.io                                                                           0.0s
------
 > [internal] load metadata for docker.io/nvidia/cuda:9.1-devel-ubuntu20.04:
------
failed to solve with frontend dockerfile.v0: failed to create LLB definition: docker.io/nvidia/cuda:9.1-devel-ubuntu20.04: not found
ERROR: Service 'ubuntu-cuda-cpp' failed to build : Build failed
Error: `docker-compose --file /home/antoine/arrow/dev/docker-compose.yml build --build-arg BUILDKIT_INLINE_CACHE=1 ubuntu-cuda-cpp` exited with a non-zero exit code 1, see the process log above.

Reporter: Antoine Pitrou / @pitrou
Assignee: Antoine Pitrou / @pitrou

Related issues:

• [C++][CI][GPU] Add CUDA CI (is depended upon by)
• [CI][Python] Install Numba on CUDA images (is depended upon by)

PRs and other links:

• GitHub Pull Request #13896

_{Note: This issue was originally created as ARROW-17423. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:
@assignUser  Do you want to take this up?

[asfimport]

Antoine Pitrou / @pitrou:
Note you probably need to install the NVidia container runtime on the host machine so that Docker guests can access the GPU.

For example on Ubuntu you should install the nvidia-container-runtime package (see https://github.com/NVIDIA/nvidia-container-runtime), then enable it in the Docker config: https://github.com/NVIDIA/nvidia-container-runtime#daemon-configuration-file

[asfimport]

Antoine Pitrou / @pitrou:
Issue resolved by pull request 13896
#13896