TL,DR

A caching proxy for Docker; allows centralised management of (multiple) registries and their authentication; caches images from any registry.

What?

Essentially, it's a man in the middle: an intercepting proxy based on nginx, to which all docker traffic is directed using the HTTPS_PROXY mechanism and injected CA root certificates.

The main feature is Docker layer/image caching, including layers served from S3, Google Storage, etc.

As a bonus it allows for centralized management of Docker registry credentials, which can in itself be the main feature, eg in Kubernetes environments.

You configure the Docker clients (err... Kubernetes Nodes?) once, and then all configuration is done on the proxy -- for this to work it requires inserting a root CA certificate into system trusted root certs.

master/:latest is unstable/beta

• :latest and :latest-debug Docker tag is unstable, built from master, and amd64-only
• Production/stable is 0.4.1, see 0.4.1 tag on Github - this image is multi-arch amd64/arm64
• The previous version is 0.3.0, see 0.3.0 tag on Github (amd64 only)

Usage

• Run the proxy on a host close (network-wise: high bandwidth, same-VPC, etc) to the Docker clients
• Expose port 3128 to the network
• Map volume /docker_mirror_cache for up to CACHE_MAX_SIZE (32gb by default) of cached images across all cached registries
• Map volume /ca, the proxy will store the CA certificate here across restarts. Important this is security sensitive.
• Env CACHE_MAX_SIZE (default 32g): set the max size to be used for caching local Docker image layers. Use Nginx sizes.
• Env REGISTRIES: space separated list of registries to cache; no need to include DockerHub, its already done internally.
• Env AUTH_REGISTRIES: space separated list of hostname:username:password authentication info.
  • hostnames listed here should be listed in the REGISTRIES environment as well, so they can be intercepted.
• Env AUTH_REGISTRIES_DELIMITER to change the separator between authentication info. By default, a space: " ". If you use keys that contain spaces (as with Google Cloud Registry), you should update this variable, e.g. setting it to AUTH_REGISTRIES_DELIMITER=";;;". In that case, AUTH_REGISTRIES could contain something like registry1.com:user1:pass1;;;registry2.com:user2:pass2.
• Env AUTH_REGISTRY_DELIMITER to change the separator between authentication info parts. By default, a colon: ":". If you use keys that contain single colons, you should update this variable, e.g. setting it to AUTH_REGISTRIES_DELIMITER=":::". In that case, AUTH_REGISTRIES could contain something like registry1.com:::user1:::pass1 registry2.com:::user2:::pass2.

Simple (no auth, all cache)

docker run --rm --name docker_registry_proxy -it \
       -p 0.0.0.0:3128:3128 \
       -v $(pwd)/docker_mirror_cache:/docker_mirror_cache \
       -v $(pwd)/docker_mirror_certs:/ca \
       rpardini/docker-registry-proxy:0.4.1

DockerHub auth

For Docker Hub authentication:

• hostname should be auth.docker.io
• username should NOT be an email, use the regular username

docker run --rm --name docker_registry_proxy -it \
       -p 0.0.0.0:3128:3128 \
       -v $(pwd)/docker_mirror_cache:/docker_mirror_cache \
       -v $(pwd)/docker_mirror_certs:/ca \
       -e REGISTRIES="k8s.gcr.io gcr.io quay.io your.own.registry another.public.registry" \
       -e AUTH_REGISTRIES="auth.docker.io:dockerhub_username:dockerhub_password your.own.registry:username:password" \
       rpardini/docker-registry-proxy:0.4.1

Simple registries auth (HTTP Basic auth)

For regular registry auth (HTTP Basic), the hostname should be the registry itself... unless your registry uses a different auth server.

See the example above for DockerHub, adapt the your.own.registry parts (in both ENVs).

This should work for quay.io also, but I have no way to test.

GitLab auth

GitLab may use a different/separate domain to handle the authentication procedure.

Just like DockerHub uses auth.docker.io, GitLab uses its primary (git) domain for the authentication.

If you run GitLab on git.example.com and its registry on reg.example.com, you need to include both in REGISTRIES and use the primary domain for AUTH_REGISTRIES.

For GitLab.com itself the authentication domain should be gitlab.com.

docker run  --rm --name docker_registry_proxy -it \
       -p 0.0.0.0:3128:3128 \
       -v $(pwd)/docker_mirror_cache:/docker_mirror_cache \
       -v $(pwd)/docker_mirror_certs:/ca \
       -e REGISTRIES="reg.example.com git.example.com" \
       -e AUTH_REGISTRIES="git.example.com:USER:PASSWORD" \
       rpardini/docker-registry-proxy:0.4.1

Google Container Registry (GCR) auth

For Google Container Registry (GCR), username should be _json_key and the password should be the contents of the service account JSON. Check out GCR docs.

The service account key is in JSON format, it contains spaces (" ") and colons (":").

To be able to use GCR you should set AUTH_REGISTRIES_DELIMITER to something different than space (e.g. AUTH_REGISTRIES_DELIMITER=";;;") and AUTH_REGISTRY_DELIMITER to something different than a single colon (e.g. AUTH_REGISTRY_DELIMITER=":::").

Example with GCR using credentials from a service account from a key file servicekey.json:

docker run --rm --name docker_registry_proxy -it \
       -p 0.0.0.0:3128:3128 \
       -v $(pwd)/docker_mirror_cache:/docker_mirror_cache \
       -v $(pwd)/docker_mirror_certs:/ca \
       -e REGISTRIES="k8s.gcr.io gcr.io quay.io your.own.registry another.public.registry" \
       -e AUTH_REGISTRIES_DELIMITER=";;;" \
       -e AUTH_REGISTRY_DELIMITER=":::" \
       -e AUTH_REGISTRIES="gcr.io:::_json_key:::$(cat servicekey.json);;;auth.docker.io:::dockerhub_username:::dockerhub_password" \
       rpardini/docker-registry-proxy:0.4.1

Configuring the Docker clients / Kubernetes nodes

Let's say you setup the proxy on host 192.168.66.72, you can then curl http://192.168.66.72:3128/ca.crt and get the proxy CA certificate.

On each Docker host that is to use the cache:

• Configure Docker proxy pointing to the caching server
• Add the caching server CA certificate to the list of system trusted roots.
• Restart dockerd

Do it all at once, tested on Ubuntu Xenial, Bionic, and Focal, all systemd based:

# Add environment vars pointing Docker to use the proxy
mkdir -p /etc/systemd/system/docker.service.d
cat << EOD > /etc/systemd/system/docker.service.d/http-proxy.conf
[Service]
Environment="HTTP_PROXY=http://192.168.66.72:3128/"
Environment="HTTPS_PROXY=http://192.168.66.72:3128/"
EOD

# Get the CA certificate from the proxy and make it a trusted root.
curl http://192.168.66.72:3128/ca.crt > /usr/share/ca-certificates/docker_registry_proxy.crt
echo "docker_registry_proxy.crt" >> /etc/ca-certificates.conf
update-ca-certificates --fresh

# Reload systemd
systemctl daemon-reload

# Restart dockerd
systemctl restart docker.service

Testing

Clear dockerd of everything not currently running: docker system prune -a -f beware

Then do, for example, docker pull k8s.gcr.io/kube-proxy-amd64:v1.10.4 and watch the logs on the caching proxy, it should list a lot of MISSes.

Then, clean again, and pull again. You should see HITs! Success.

Do the same for docker pull ubuntu and rejoice.

Test your own registry caching and authentication the same way; you don't need docker login, or .docker/config.json anymore.

Developing/Debugging

Since 0.4.1 there is a separate -debug version of the image, which includes nginx-debug, and has mitmproxy (actually mitmweb) inserted after the CONNECT proxy but before the caching logic. This allows very in-depth debugging, but tends to be unstable with huge layers. Use sparingly, and definitely not in production.

docker run --rm --name docker_registry_proxy -it 
       -e DEBUG_NGINX=true -e DEBUG=true -p 0.0.0.0:8081:8081 \
       -p 0.0.0.0:3128:3128 \
       -v $(pwd)/docker_mirror_cache:/docker_mirror_cache \
       -v $(pwd)/docker_mirror_certs:/ca \
       rpardini/docker-registry-proxy:0.4.1-debug

• DEBUG=true enables the mitmweb proxy, accessible on port 8081
• DEBUG_NGINX=true enables nginx-debug and debug logging, which probably is too much.

Gotchas

• If you authenticate to a private registry and pull through the proxy, those images will be served to any client that can reach the proxy, even without authentication. beware
• Repeat, this will make your private images very public if you're not careful.
• Currently you cannot push images while using the proxy which is a shame. PRs welcome.
• Setting this on Linux is relatively easy.
  • On Mac and Windows the CA-certificate part will be very different but should work in principle.
  • Please send PRs with instructions for Windows and Mac if you succeed!

Why not use Docker's own registry, which has a mirror feature?

Yes, Docker offers Registry as a pull through cache, unfortunately it only covers the DockerHub case. It won't cache images from quay.io, k8s.gcr.io, gcr.io, or any such, including any private registries.

That means that your shiny new Kubernetes cluster is now a bandwidth hog, since every image will be pulled from the Internet on every Node it runs on, with no reuse.

This is due to the way the Docker "client" implements --registry-mirror, it only ever contacts mirrors for images with no repository reference (eg, from DockerHub). When a repository is specified dockerd goes directly there, via HTTPS (and also via HTTP if included in a --insecure-registry list), thus completely ignoring the configured mirror.

Docker itself should provide this.

Yeah. Docker Inc should do it. So should NPM, Inc. Wonder why they don't. 😼

TODO:

• Test and make auth work with quay.io, unfortunately I don't have access to it (hint, hint, quay)
• Hide the mitmproxy building code under a Docker build ARG.
• "Developer Office" proxy scenario, where many developers on a fast LAN share a proxy for bandwidth and speed savings (already works for pulls, but messes up pushes, which developers tend to use a lot)