AWS EB docker-compose deployment from private registry access forbidden - amazon-web-services

Question

I'm trying to get docker-compose deployment to AWS Elastic Beanstalk working, in which the docker images are pulled from a private registry hosted by GitLab.
The strange thing is that initial deployment works perfectly; It pulls the image from the private registry and starts the containers using docker-compose, and the webpage (served by Django) is accessible through the host.
Deploying a new version using the same docker-compose and the same docker image will result in an error while pulling the docker image:
2021/03/16 09:28:34.957094 [ERROR] An error occurred during execution of command [app-deploy] - [Run Docker Container]. Stop running the command. Error: failed to run docker containers: Command /bin/sh -c docker-compose up -d failed with error exit status 1. Stderr:Building with native build. Learn about native build in Compose here: https://docs.docker.com/go/compose-native-build/
Creating network "current_default" with the default driver
Pulling redis (redis:alpine)...
Pulling mysql (mysql:5.7)...
Pulling project.dockertest(registry.gitlab.com/company/spikes/dockertest:latest)...
Get https://registry.gitlab.com/v2/company/spikes/dockertest/manifests/latest: denied: access forbidden
2021/03/16 09:28:34.957104 [INFO] Executing cleanup logic
Setup
AWS Elastic Beanstalk 64bit Amazon Linux 2/3.2
Gitlab registry credentials are stored within a S3 bucket, with the filename .dockercfg and has the following content:
{
"auths": {
"registry.gitlab.com": {
"auth": "base64 encoded username:personal_access_token"
}
},
"HttpHeaders": {
"User-Agent": "Docker-Client/18.03.1-ce (linux)"
}
}
The repository contains a v3 Dockerrun.aws.json file to refer to the credential file in S3:
{
"AWSEBDockerrunVersion": "3",
"Authentication": {
"bucket": "gitlab-dockercfg",
"key": ".dockercfg"
}
}
Reproduce
Setup docker-compose.yml that uses a service with a private docker image (and can be pulled with the credentials setup in the dockercfg within S3)
Create a new applicatoin that uses the docker-platform.
eb init testapplication --platform=docker --region=eu-west-1
Note: region must be the same as the S3 bucket containing the dockercfg.
Initial deployment (this will succeed)
eb create testapplication-test --branch_default --cname testapplication-test --elb-type=application --instance-types=t2.micro --min-instance=1 --max-instances=4
The initial deployment shows that the image is available and can be started:
2021/03/16 08:58:07.533988 [INFO] save docker tag command: docker tag 5812dfe24a4f redis:alpine
2021/03/16 08:58:07.533993 [INFO] save docker tag command: docker tag f8fcde8b9ae2 mysql:5.7
2021/03/16 08:58:07.533998 [INFO] save docker tag command: docker tag 1dd9b65d6a9f registry.gitlab.com/company/spikes/dockertest:latest
2021/03/16 08:58:07.534010 [INFO] Running command /bin/sh -c docker rm `docker ps -aq`
Without changing anything to the local repository and the remote docker image on the private registry, lets do a redeployment which will trigger the error:
eb deploy testapplication-test
This will fail with the following output:
...
2021-03-16 10:02:28 INFO Command execution completed on all instances. Summary: [Successful: 0, Failed: 1].
2021-03-16 10:02:29 ERROR Unsuccessful command execution on instance id(s) 'i-0dc445d118ac14b80'. Aborting the operation.
2021-03-16 10:02:29 ERROR Failed to deploy application.
ERROR: ServiceError - Failed to deploy application.
And logs of the instance show (/var/log/eb-engine.log):
Pulling redis (redis:alpine)...
Pulling mysql (mysql:5.7)...
Pulling project.dockertest (registry.gitlab.com/company/spikes/dockertest:latest)...
Get https://registry.gitlab.com/v2/company/spikes/dockertest/manifests/latest: denied: access forbidden
2021/03/16 10:02:25.902479 [INFO] Executing cleanup logic
Steps I've tried to debug or solve the issue
Rename dockercfg to .dockercfg on S3 (somewhere mentioned on the internet as possible solution)
Use the 'old' docker config format instead of the one generated by docker 1.7+. But later on I figured out that Amazon Linux 2-instances are compatible with the new format together with Dockerrun v3
Having an incorrectly formatted dockercfg on S3 will cause an error deployment regarding the misformatted file (so it actually does something with the dockercfg from S3)
Documentation
https://docs.aws.amazon.com/elasticbeanstalk/latest/dg/single-container-docker-configuration.html
I'm out of debug options, and I've no idea where to look any further to debug this problem. Perhaps someone can see what is going wrong here?

Answer 1

First of all, the issue describe above is a bug confirmed by Amazon. To get the deployment working on our side, we've contacted Amazon support.
They've a fix in place which should be released this month, so keep an eye on the changelog of the Elastic beanstalk platform: https://docs.aws.amazon.com/elasticbeanstalk/latest/relnotes/relnotes.html
Although the upcoming release should have the fix, there is a workaround available to get the docker-compose deployment working.
Elastic Beanstalk allows hook to be executed within the deployment, which can be used to fetch the .docker.cfg from a S3 bucket to authenticate with against the private registry.
To do so, create the following file and directories from the root of the project:
File location: .platform/hooks/predeploy/docker_login
#!/bin/bash
aws s3 cp s3://{{bucket_name_to_use}}/.dockercfg ~/.docker/config.json
Important: Add execution rights to this file (for example: chmod +x .platform/hooks/predeploy/docker_login)
To support instance configuration changes, please symlink the hooks directory to confighooks:
ln -s .platform/hooks/ .platform/confighooks/
Updating configuration requires the .dockercfg credentials to be fetched too.
This should enable continuous deployments to the same EB-instance without the authentication errors, because the hook will be execute before the docker image pulling.
Some background:
The docker daemon reads credentials from ~/.docker/config by default on traditional linux systems. On the initial deploy this file will exist on the Elastic Beanstalk instance. On the next deployment this file is removed. Unfortunately, on the next deployment the .dockercfg is not refetched, therefor the docker daemon does not have the correct credentials to authenticate with.

Answer 2

I was dealing the same errors while trying to pull images from a privately hosted GitLab instance. I was able to resolve them by including the email address that was associated with the generated token found in the auth field of the .dockercfg file.
The following file format worked for me:
"registry.gitlab.com" {
"auth": "base64 encoded username:personal_access_token",
"email": "email for personal access token"
}
In my case I used a Project Access Token, which has an e-mail address associated with it once it is created.
The file format in the Elastic Beanstalk documentation for the authentication file here, indicates that this is the required file format, though the versions that it says this format is required for are almost certainly outdated, since we are running Docker ^19.