Should "run" command ignore restart policy?

[stefcl]

Hello,
I am using docker-compose 1.6.2 along with docker 1.10.3.
I often run short-lived containers in interactive mode using docker-compose run with --rm option and an argument line for debugging purpose before running them in background.

Sometimes, it takes many attempts to get the command line of my dockerized apps right. Recently, I found a lot of dangling containers with suffixes like "run_3" "run_4" while running docker ps -a, these seemed to correspond to failed "run" attempts that the docker engine was permanently trying to "resurrect". It comes from my yml file which sets the restart policy to "unless-stopped" for all my containers.

While I understand why it happens, I am wondering if the docker-compose run command should not ignore the restart policy defined in the yml file. The run command seems to be designed primarily for interactive testing and ephemeral containers, while restart policies make most sense with long-lived services.
Kind regards

[aanand]

The restart policy should be ignored if you pass --rm. If it's not, could you provide the exact commands you ran to reproduce the problem?

[stefcl]

I am probably the one at fault for omitting the --rm flag since the docker daemon could not attempt to restart a container which has been removed.
However, when you don't pass --rm, docker applies the restart policy, altough the terminal is no longer attached. I can't think of any use case in which one would use run non-interactively and still want auto-restart, but the current behavior is quite reasonable. I guess you can close this issue.

[aanand]

I'm also not sure there's any valid use case for run with a restart policy. The reason we went with only ignoring it for --rm was parity with docker run.

I can imagine there are people out there using docker-compose run -d web to start their app, who'll be surprised if the restart policy isn't honoured... but they should be using docker-compose up -d web.

[stefcl]

Given that there are already some exceptions, especially port mappings, that do not make run a really good choice for production containers, I suppose it would not hurt that much to add a special rule with the restart directive, but I'd understand if you prefer to leave the current state unchanged to avoid introducing too many special implicit rules.
If you don't use --rm, you are probably interested in the resulting container, so you'll probably notice that it's stuck in a restart loop.