Dependency Proxy
The Dependency Proxy is a pull-through-cache for public registry images from DockerHub. This document describes how this feature is constructed in GitLab.
Container registry
The Dependency Proxy for the container registry acts a stand-in for a remote container registry. In our case, the remote registry is the public DockerHub registry.
From the user’s perspective, the GitLab instance is just a container registry that they are interacting with to
pull images by using docker login gitlab.com
When you use docker login gitlab.com, the Docker client uses the v2 API
to make requests.
To support authentication, we must include one route:
To support docker pull requests, we must include two additional routes:
These routes are defined in gitlab-org/gitlab/config/routes/group.rb.
In its simplest form, the Dependency Proxy manages three requests:
- Logging in / returning a JWT
- Fetching a manifest
- Fetching a blob
Here is what the general request sequence looks like for the Dependency Proxy:
Authentication and authorization
When a Docker client authenticates with a registry, the registry tells the client where to get a JSON Web Token
(JWT) and to use it for all subsequent requests. This allows the authentication service to live in a separate
application from the registry. For example, the GitLab container registry directs Docker clients to get a token
from https://gitlab.com/jwt/auth. This endpoint is part of the gitlab-org/gitlab project, also known as the
rails project or web service.
When a user tries to sign in to the dependency proxy with a Docker client, we must tell it where to get a JWT. We
can use the same endpoint we use with the container registry: https://gitlab.com/jwt/auth. But in our case,
we tell the Docker client to specify service=dependency_proxy in the parameters so can use a separate underlying
service to generate the token.
This sequence diagram shows the request flow for logging into the Dependency Proxy.
The dependency proxy uses its own authentication service, separate from the authentication managed by the UI
(ApplicationController) and API (ApiGuard). Once the service has created a JWT, the DependencyProxy::ApplicationController
manages authentication and authorization for the rest of the requests. It manages the user by using GitLab::Auth::Result and
is similar to the authentication implemented by the Git client requests in GitHttpClientController.
Caching
Blobs are cached artifacts with no logic around them. We cache them by digest. When we receive a request for a new blob, we check to see if we have a blob with the requested digest, and return it. Otherwise we fetch it from the external registry and cache it.
Manifests are more complicated, partially due to rate limiting on DockerHub.
A manifest is essentially a recipe for creating an image. It has a list of blobs to create a certain image. So
alpine:latest has a manifest associated with it that specifies the blobs needed to create the alpine:latest
image. The interesting part is that alpine:latest can change over time, so we can’t just cache the manifest and
assume it is OK to use forever. Instead, we must check the digest of the manifest, which is an ETag. This gets
interesting because the requests for manifests often don’t include the digest. So how do we know if the manifest
we have cached is still the most up-to-date alpine:latest? DockerHub allows free HEAD requests that don’t count
toward their rate limit. The HEAD request returns the manifest digest so we can tell whether or not the one we
have is stale.
With this knowledge, we have built the following logic to manage manifest requests:
Workhorse for file handling
Management of file uploads and caching happens in Workhorse. This explains the additional
POST routes
that we have for the Dependency Proxy.
The send_dependency
method makes a request to Workhorse including the previously fetched JWT from the external registry. Workhorse then
can use that token to request the manifest or blob the user originally requested. The Workhorse code lives in
workhorse/internal/dependencyproxy/dependencyproxy.go.
Once we put it all together, the sequence for requesting an image file looks like this:
Cleanup policies
The cleanup policies for the Dependency Proxy work as time-to-live policies. They allow users to set the number
of days a file is allowed to remain cached if it has been unread. Since there is no way to associate the blobs
with the images they belong to (to do this, we would need to build the metadata database that the container registry
folks built), we can set up rules like “if this blob has not been pulled in 90 days, delete it”. This means that
any files that are continuously getting pulled will not be removed from the cache, but if, for example,
alpine:latest changes and one of the underlying blobs is no longer used, it will eventually get cleaned up
because it has stopped getting pulled. We use the read_at attribute to track the last time a given
dependency_proxy_blob or dependency_proxy_manifest was pulled.
These work using a cron worker, DependencyProxy::CleanupDependencyProxyWorker, that will kick off two limited capacity workers: one to delete blobs, and one to delete manifests. The capacity is set in an application setting.
Historic reference links
- Dependency proxy for private groups - initial authentication implementation
- Manifest caching - initial manifest caching implementation
- Workhorse for blobs - initial workhorse implementation
- Workhorse for manifest - moving manifest cache logic to Workhorse
- Deploy token support - authorization largely updated
- SSO support - changes how policies are checked