linux - `docker run` on a remote host

Question

is it possible (using the docker command or the docker-py API directly) to start a container from a remote host?

Lets assume I have two machines that have different architectures: - A is an x86 machine - B is an ARM machine

I would like to run a container on the B machine using my A machine. At first, I thought it was possible using this command:

[A]$> DOCKER_HOST=$MACHINE_B_IP:$MACHIN_B_PORT docker run hello-from-B

But this command actually pulls the image hello-from-B and tries to run it on the machine A which ends up on some exec format error cause obviously you can't run images that are specific to ARM to an x86 machine.

Communication between machine A and B is working well. I can run commands like images or ps and it gives me the expected results:

[A]$> DOCKER_HOST=$MACHINE_B_IP:$MACHIN_B_PORT docker images
REPOSITORY                TAG                 IMAGE ID            CREATED             SIZE
hello-from-B              <none>              fd5059044831        13 hours ago        1.26GB

I've heard about docker-machine and haven't tried it yet, but from my understanding, this won't solve my problem.

Is there any way to achieve that using docker directly. A workaround might be to using ssh to connect to the remote host and use the docker client directly from the remote host, but I'd like to avoid this solution as much as possible.

Thanks in advance,

---

TL;DR;

How can DOCKER_HOST=... docker run something runs something on the DOCKER_HOST rather than running it on my local machine.

Answer 1

Check if the latest docker 18.09 includes that feature.
See docker/cli PR 1014

Added support for SSH connection. e.g. docker -H ssh://me@server

• The cli should accept ssh://me@server for DOCKER_HOST and -H. Using that would execute ssh with the passed config.
• The ssh command would call a hidden command on the docker CLI binary on the remote side. For example, docker dial-stdio.

This command will make a connection to the local DOCKER_HOST variable (almost always the default local socket) and forward that connection on the commands stdio.
Even though this command is supposed to run locally to the dockerd binary, we think that it is an invalid configuration for this feature to remove the local docker binary so we can rely on it always being present.

How to verify it

docker -H ssh://me@server run -it --rm busybox

The reaction so far:

From ops and sysadmins everywhere, we thank you for this fantastic and unexpected feature.
I'm hoping this will seriously cut down the number of times I see people opening dockerd TCP w/o TLS and just opt for SSH endpoints for remote mgmt.