In order to test an nginx docker in CircleCI before pushing it on my register, I use dgoss (which use docker commands) and fake some internal DNS by overwriting the /etc/hosts file like this:
version: 2
workflows:
version: 2
build_deploy:
jobs:
- deploy_on_registery
jobs:
deploy_on_registery:
docker:
- image: node:lts
steps:
- checkout
- setup_remote_docker
- run:
name: Dockerize project
command: docker build -t someapp:test .
- run:
name: Install Goss (https://github.com/aelsabbahy/goss)
command: curl -fsSL https://goss.rocks/install | sh
- run:
name: Install test dependency
command: npm install -g node-static
- run:
name: Fake server
background: true
command: |
static -p 80
- run:
name: Fake server
background: true
command: |
static -p 8000
- run:
name: Configure network
command: |
echo "127.0.0.1 someinternaldns" >> /etc/hosts
echo "127.0.0.1 someotherinternaldns" >> /etc/hosts
- run:
name: Test Docker image
command: |
curl http://someinternaldns # this works
GOSS_FILES_STRATEGY=cp dgoss run -v $PWD/secret/:/secret/ someapp:test # this doesn't work and returns `nginx: [emerg] host not found in upstream "someinternaldns" in /etc/nginx/conf.d/default.conf:171`
It worked perfectly on my computer but it doesn’t on CircleCI. I got an error which said that the host doesn’t exist like if the /etc/hosts wasn’t used by the dgoss. Isn’t the /etc/hosts shared between the runner and the docker commands inside it?
I see that the instruction to be able to use docker commands (setup_remote_docker) is quite explicit about the fact that it is a remote (unlike my local docker). So my question is: is it a bug? a missing feature? or if it is desired, how can I modify my /etc/hosts or simply add my internal DNS without altering my docker?
In general, Docker containers do not like their /etc/hosts being written to. The way to do this, at least with the docker run command, is to use --add-host. That might work here, though it depends on whether your dgoss allows it.
It also depends on whether the underlying Docker platform gives you enough permissions to modify networking features. For example, you cannot publish ports in CircleCI to the build host, I think because if you could do that, you would have enough control over the networking stack to affect other customers on the same bare-metal machine.
My way around that is to use Docker Compose, which works very well indeed. I am pretty sure the DC YAML config format allows extra hosts to be created, so if docker switches do not work, try moving to DC. I have not tried this in CircleCI, but it is what I would try.
The other thing you can try is to move to a Machine executor. This would be one layer of Docker, instead of Docker-in-Docker, and you would be able to modify your hosts file there. However you get less control over your build image - it is not Docker any more - and technically the charging model has not been finalised for it (the docs note that it could get more expensive).
