Troubleshooting the container registry

Before diving in to the following sections, here's some basic troubleshooting:

1. Check to make sure that the system clock on your Docker client and GitLab server have been synchronized (for example, via NTP).

2. If you are using an S3-backed Registry, double check that the IAM permissions and the S3 credentials (including region) are correct. See the sample IAM policy for more details.

3. Check the Registry logs (for example /var/log/gitlab/registry/current) and the GitLab production logs for errors (for example /var/log/gitlab/gitlab-rails/production.log). You may be able to find clues there.

Using self-signed certificates with container registry

If you're using a self-signed certificate with your container registry, you might encounter issues during the CI jobs like the following:

Error response from daemon: Get registry.example.com/v1/users/: x509: certificate signed by unknown authority

The Docker daemon running the command expects a cert signed by a recognized CA, thus the error above.

While GitLab doesn't support using self-signed certificates with Container Registry out of the box, it is possible to make it work by instructing the Docker daemon to trust the self-signed certificates, mounting the Docker daemon and setting privileged = false in the GitLab Runner config.toml file. Setting privileged = true takes precedence over the Docker daemon:

  [runners.docker]
    image = "ruby:2.6"
    privileged = false
    volumes = ["/var/run/docker.sock:/var/run/docker.sock", "/cache"]

Additional information about this: issue 18239.

Docker login attempt fails with: 'token signed by untrusted key'

Registry relies on GitLab to validate credentials If the registry fails to authenticate valid login attempts, you get the following error message:

# docker login gitlab.company.com:4567
Username: user
Password:
Error response from daemon: login attempt to https://gitlab.company.com:4567/v2/ failed with status: 401 Unauthorized

And more specifically, this appears in the /var/log/gitlab/registry/current log file:

level=info msg="token signed by untrusted key with ID: "TOKE:NL6Q:7PW6:EXAM:PLET:OKEN:BG27:RCIB:D2S3:EXAM:PLET:OKEN""
level=warning msg="error authorizing context: invalid token" go.version=go1.12.7 http.request.host="gitlab.company.com:4567" http.request.id=74613829-2655-4f96-8991-1c9fe33869b8 http.request.method=GET http.request.remoteaddr=10.72.11.20 http.request.uri="/v2/" http.request.useragent="docker/19.03.2 go/go1.12.8 git-commit/6a30dfc kernel/3.10.0-693.2.2.el7.x86_64 os/linux arch/amd64 UpstreamClient(Docker-Client/19.03.2 \(linux\))"

GitLab uses the contents of the certificate key pair's two sides to encrypt the authentication token for the Registry. This message means that those contents do not align.

Check which files are in use:

• grep -A6 'auth:' /var/opt/gitlab/registry/config.yml

  ## Container registry certificate
     auth:
       token:
         realm: https://gitlab.my.net/jwt/auth
         service: container_registry
         issuer: omnibus-gitlab-issuer
    -->  rootcertbundle: /var/opt/gitlab/registry/gitlab-registry.crt
         autoredirect: false

• grep -A9 'Container Registry' /var/opt/gitlab/gitlab-rails/etc/gitlab.yml

  ## Container registry key
     registry:
       enabled: true
       host: gitlab.company.com
       port: 4567
       api_url: http://127.0.0.1:5000 # internal address to the registry, is used by GitLab to directly communicate with API
       path: /var/opt/gitlab/gitlab-rails/shared/registry
  -->  key: /var/opt/gitlab/gitlab-rails/etc/gitlab-registry.key
       issuer: omnibus-gitlab-issuer
       notification_secret:

The output of these openssl commands should match, proving that the cert-key pair is a match:

/opt/gitlab/embedded/bin/openssl x509 -noout -modulus -in /var/opt/gitlab/registry/gitlab-registry.crt | /opt/gitlab/embedded/bin/openssl sha256
/opt/gitlab/embedded/bin/openssl rsa -noout -modulus -in /var/opt/gitlab/gitlab-rails/etc/gitlab-registry.key | /opt/gitlab/embedded/bin/openssl sha256

If the two pieces of the certificate do not align, remove the files and run gitlab-ctl reconfigure to regenerate the pair. The pair is recreated using the existing values in /etc/gitlab/gitlab-secrets.json if they exist. To generate a new pair, delete the registry section in your /etc/gitlab/gitlab-secrets.json before running gitlab-ctl reconfigure.

If you have overridden the automatically generated self-signed pair with your own certificates and have made sure that their contents align, you can delete the 'registry' section in your /etc/gitlab/gitlab-secrets.json and run gitlab-ctl reconfigure.

AWS S3 with the GitLab registry error when pushing large images

When using AWS S3 with the GitLab registry, an error may occur when pushing large images. Look in the Registry log for the following error:

level=error msg="response completed with error" err.code=unknown err.detail="unexpected EOF" err.message="unknown error"

To resolve the error specify a chunksize value in the Registry configuration. Start with a value between 25000000 (25 MB) and 50000000 (50 MB).

::Tabs

:::TabTitle Linux package (Omnibus)

1. Edit /etc/gitlab/gitlab.rb:

   registry['storage'] = {
     's3' => {
       'accesskey' => 'AKIAKIAKI',
       'secretkey' => 'secret123',
       'bucket'    => 'gitlab-registry-bucket-AKIAKIAKI',
       'chunksize' => 25000000
     }
   }

2. Save the file and reconfigure GitLab for the changes to take effect.

:::TabTitle Self-compiled (source)

1. Edit config/gitlab.yml:

   storage:
     s3:
       accesskey: 'AKIAKIAKI'
       secretkey: 'secret123'
       bucket: 'gitlab-registry-bucket-AKIAKIAKI'
       chunksize: 25000000

2. Save the file and restart GitLab for the changes to take effect.

::EndTabs

Supporting older Docker clients

The Docker container registry shipped with GitLab disables the schema1 manifest by default. If you are still using older Docker clients (1.9 or older), you may experience an error pushing images. See issue 4145 for more details.

You can add a configuration option for backwards compatibility.

::Tabs

:::TabTitle Linux package (Omnibus)

1. Edit /etc/gitlab/gitlab.rb:

   registry['compatibility_schema1_enabled'] = true

2. Save the file and reconfigure GitLab for the changes to take effect.

:::TabTitle Self-compiled (source)

1. Edit the YAML configuration file you created when you deployed the registry. Add the following snippet:

   compatibility:
       schema1:
           enabled: true

2. Restart the registry for the changes to take affect.

::EndTabs

Docker connection error

A Docker connection error can occur when there are special characters in either the group, project or branch name. Special characters can include:

• Leading underscore
• Trailing hyphen/dash
• Double hyphen/dash

To get around this, you can change the group path, change the project path or change the branch name. Another option is to create a push rule to prevent this error for the entire instance.

Image push errors

You might get stuck in retry loops when pushing Docker images, even though docker login succeeds.

This issue occurs when NGINX isn't properly forwarding headers to the registry, typically in custom setups where SSL is offloaded to a third-party reverse proxy.

For more information, see Docker push through NGINX proxy fails trying to send a 32B layer #970.

To resolve this issue, update your NGINX configuration to enable relative URLs in the registry:

::Tabs

:::TabTitle Linux package (Omnibus)

1. Edit /etc/gitlab/gitlab.rb:

   registry['env'] = {
     "REGISTRY_HTTP_RELATIVEURLS" => true
   }

2. Save the file and reconfigure GitLab for the changes to take effect.

:::TabTitle Self-compiled (source)

1. Edit the YAML configuration file you created when you deployed the registry. Add the following snippet:

   http:
       relativeurls: true

2. Save the file and restart GitLab for the changes to take effect.

:::TabTitle Docker Compose

1. Edit your docker-compose.yaml file:

   GITLAB_OMNIBUS_CONFIG: |
     registry['env'] = {
       "REGISTRY_HTTP_RELATIVEURLS" => true
     }

2. If the issue persists, ensure both URLs use HTTPS:

   GITLAB_OMNIBUS_CONFIG: |
     external_url 'https://git.example.com'
     registry_external_url 'https://git.example.com:5050'

3. Save the file and restart the container:

   sudo docker restart gitlab

::EndTabs

Enable the Registry debug server

You can use the container registry debug server to diagnose problems. The debug endpoint can monitor metrics and health, as well as do profiling.

WARNING: Sensitive information may be available from the debug endpoint. Access to the debug endpoint must be locked down in a production environment.

The optional debug server can be enabled by setting the registry debug address in your gitlab.rb configuration.

registry['debug_addr'] = "localhost:5001"

After adding the setting, reconfigure GitLab to apply the change.

Use curl to request debug output from the debug server:

curl "localhost:5001/debug/health"
curl "localhost:5001/debug/vars"

Enable Registry Prometheus Metrics

If the debug server is enabled, you can also enable Prometheus metrics. This endpoint exposes highly detailed telemetry related to almost all registry operations.

registry['debug'] = {
  'prometheus' => {
    'enabled' => true,
    'path' => '/metrics'
  }
}

Use curl to request debug output from Prometheus:

curl "localhost:5001/debug/metrics"

Tags with an empty name

If using AWS DataSync to copy the registry data to or between S3 buckets, an empty metadata object is created in the root path of each container repository in the destination bucket. This causes the registry to interpret such files as a tag that appears with no name in the GitLab UI and API. For more information, see this issue.

To fix this you can do one of two things:

• Use the AWS CLI rm command to remove the empty objects from the root of each affected repository. Pay special attention to the trailing / and make sure not to use the --recursive option:

  aws s3 rm s3://<bucket>/docker/registry/v2/repositories/<path to repository>/

• Use the AWS CLI sync command to copy the registry data to a new bucket and configure the registry to use it. This leaves the empty objects behind.

Advanced Troubleshooting

We use a concrete example to illustrate how to diagnose a problem with the S3 setup.

Investigate a cleanup policy

If you're unsure why your cleanup policy did or didn't delete a tag, execute the policy line by line by running the below script from the Rails console. This can help diagnose problems with the policy.

repo = ContainerRepository.find(<repository_id>)
policy = repo.project.container_expiration_policy

tags = repo.tags
tags.map(&:name)

tags.reject!(&:latest?)
tags.map(&:name)

regex_delete = ::Gitlab::UntrustedRegexp.new("\\A#{policy.name_regex}\\z")
regex_retain = ::Gitlab::UntrustedRegexp.new("\\A#{policy.name_regex_keep}\\z")

tags.select! { |tag| regex_delete.match?(tag.name) && !regex_retain.match?(tag.name) }

tags.map(&:name)

now = DateTime.current
tags.sort_by! { |tag| tag.created_at || now }.reverse! # Lengthy operation

tags = tags.drop(policy.keep_n)
tags.map(&:name)

older_than_timestamp = ChronicDuration.parse(policy.older_than).seconds.ago

tags.select! { |tag| tag.created_at && tag.created_at < older_than_timestamp }

tags.map(&:name)

• The script builds the list of tags to delete (tags).
• tags.map(&:name) prints a list of tags to remove. This may be a lengthy operation.
• After each filter, check the list of tags to see if it contains the intended tags to destroy.

Unexpected 403 error during push

A user attempted to enable an S3-backed Registry. The docker login step went fine. However, when pushing an image, the output showed:

The push refers to a repository [s3-testing.myregistry.com:5050/root/docker-test/docker-image]
dc5e59c14160: Pushing [==================================================>] 14.85 kB
03c20c1a019a: Pushing [==================================================>] 2.048 kB
a08f14ef632e: Pushing [==================================================>] 2.048 kB
228950524c88: Pushing 2.048 kB
6a8ecde4cc03: Pushing [==>                                                ] 9.901 MB/205.7 MB
5f70bf18a086: Pushing 1.024 kB
737f40e80b7f: Waiting
82b57dbc5385: Waiting
19429b698a22: Waiting
9436069b92a3: Waiting
error parsing HTTP 403 response body: unexpected end of JSON input: ""

This error is ambiguous, as it's not clear whether the 403 is coming from the GitLab Rails application, the Docker Registry, or something else. In this case, since we know that since the login succeeded, we probably need to look at the communication between the client and the Registry.

The REST API between the Docker client and Registry is described in the Docker documentation. Usually, one would just use Wireshark or tcpdump to capture the traffic and see where things went wrong. However, since all communications between Docker clients and servers are done over HTTPS, it's a bit difficult to decrypt the traffic quickly even if you know the private key. What can we do instead?

One way would be to disable HTTPS by setting up an insecure Registry. This could introduce a security hole and is only recommended for local testing. If you have a production system and can't or don't want to do this, there is another way: use mitmproxy, which stands for Man-in-the-Middle Proxy.

mitmproxy

mitmproxy allows you to place a proxy between your client and server to inspect all traffic. One wrinkle is that your system needs to trust the mitmproxy SSL certificates for this to work.

The following installation instructions assume you are running Ubuntu:

1. Install mitmproxy.

2. Run mitmproxy --port 9000 to generate its certificates. Enter CTRL-C to quit.

3. Install the certificate from ~/.mitmproxy to your system:

   sudo cp ~/.mitmproxy/mitmproxy-ca-cert.pem /usr/local/share/ca-certificates/mitmproxy-ca-cert.crt
   sudo update-ca-certificates

If successful, the output should indicate that a certificate was added:

Updating certificates in /etc/ssl/certs... 1 added, 0 removed; done.
Running hooks in /etc/ca-certificates/update.d....done.

To verify that the certificates are properly installed, run:

mitmproxy --port 9000

This command runs mitmproxy on port 9000. In another window, run:

curl --proxy "http://localhost:9000" "https://httpbin.org/status/200"

If everything is set up correctly, information is displayed on the mitmproxy window and no errors are generated by the curl commands.

Running the Docker daemon with a proxy

For Docker to connect through a proxy, you must start the Docker daemon with the proper environment variables. The easiest way is to shutdown Docker (for example sudo initctl stop docker) and then run Docker by hand. As root, run:

export HTTP_PROXY="http://localhost:9000"
export HTTPS_PROXY="http://localhost:9000"
docker daemon --debug

This command launches the Docker daemon and proxies all connections through mitmproxy.

Running the Docker client

Now that we have mitmproxy and Docker running, we can attempt to sign in and push a container image. You may need to run as root to do this. For example:

docker login s3-testing.myregistry.com:5050
docker push s3-testing.myregistry.com:5050/root/docker-test/docker-image

In the example above, we see the following trace on the mitmproxy window:

The above image shows:

• The initial PUT requests went through fine with a 201 status code.
• The 201 redirected the client to the S3 bucket.
• The HEAD request to the AWS bucket reported a 403 Unauthorized.

What does this mean? This strongly suggests that the S3 user does not have the right permissions to perform a HEAD request. The solution: check the IAM permissions again. Once the right permissions were set, the error goes away.

Missing gitlab-registry.key prevents container repository deletion

If you disable your GitLab instance's container registry and try to remove a project that has container repositories, the following error occurs:

Errno::ENOENT: No such file or directory @ rb_sysopen - /var/opt/gitlab/gitlab-rails/etc/gitlab-registry.key

In this case, follow these steps:

1. Temporarily enable the instance-wide setting for the container registry in your gitlab.rb:

   gitlab_rails['registry_enabled'] = true

2. Save the file and reconfigure GitLab for the changes to take effect.

3. Try the removal again.

If you still can't remove the repository using the common methods, you can use the GitLab Rails console to remove the project by force:

# Path to the project you'd like to remove
prj = Project.find_by_full_path(<project_path>)

# The following will delete the project's container registry, so be sure to double-check the path beforehand!
if prj.has_container_registry_tags?
  prj.container_repositories.each { |p| p.destroy }
end

Registry service listens on IPv6 address instead of IPv4

You might see the following error if the localhost hostname resolves to a IPv6 loopback address (::1) on your GitLab server and GitLab expects the registry service to be available on the IPv4 loopback address (127.0.0.1):

request: "GET /v2/ HTTP/1.1", upstream: "http://[::1]:5000/v2/", host: "registry.example.com:5005"
[error] 1201#0: *13442797 connect() failed (111: Connection refused) while connecting to upstream, client: x.x.x.x, server: registry.example.com, request: "GET /v2/<path> HTTP/1.1", upstream: "http://[::1]:5000/v2/<path>", host: "registry.example.com:5005"

To fix the error, change registry['registry_http_addr'] to an IPv4 address in /etc/gitlab/gitlab.rb. For example:

registry['registry_http_addr'] = "127.0.0.1:5000"

See issue 5449 for more details.

Push failures and high CPU usage with Google Cloud Storage (GCS)

You might get a 502 Bad Gateway error when pushing container images to a registry that uses GCS as the backend. The registry might also experience CPU usage spikes when pushing large images.

This issue occurs when the registry communicates with GCS using the HTTP/2 protocol.

The workaround is to disable HTTP/2 in your registry deployment by setting the GODEBUG environment variable to http2client=0.

For more information, see issue 1425.