cartorio is meant to act as a thin layer on top of content to serve that content from consumers of the registry API.


                      registry
                        API
                         |
 ---------.              |
          |              |
 content  +--- cartorio -+-->  docker pull
          |              |
 ---------*              |
                         |

Here you can find an answer to “what’s the minimum I should implement to serve images to container engines?", or, how does that REGISTRY API piece looks like in practice.

ps.: this article assumes that you’re familiar with the basic usage of Docker (creating containers, images, etc).

literature

ps.: if you don’t care about specifications and standards, go straight to the next section

Most of the container-related tech that we use is now in the process of being standardized under the Open Containers Initiative (OCI):

For the purpose of describing cartorio, here we focus only on the later - the distribution spec - even though pratically, that’s pretty much the same as the Docker Registry HTTP API V2.

following the pcap

What happens “on the wire” when you docker pull $image?


    REGISTRY                    DOCKERD


                |        |
  image1        |   ???  |
  image2    <---+--------+----> docker pull
  ....          |        |
  imageN        |        |
                |        |

Capturing the request flow, we can see the entirety of what’s necessary for having container images distributed through a registry.

To get started, let’s create an image from a Dockerfile and see how what happens when we pull it from a registry.


a simple image

One of the most simple types of images that we could create is one that just adds a file to it.

# starting with a completely empty layer.
#
FROM scratch              

# add a filke that we have locally
#
ADD ./file.txt /file.txt

Build this image, and we can see the layers generated:

# build the Dockerfile under `./assets`, having the context
# from `assets`, then tag the final layer as `file` (this could
# be something like `concourse/concourse`.
#
docker build --tag file ./assets


docker history file
IMAGE           CREATED BY                           
24cca6f78bbd    ADD ./file.txt /file.txt # buildkit 

Having the image there, we can now push it to a registry and inspect the request flow.


“sniffing the wire”

Looking at the result from capturing the packets from a docker pull (i.e., putting ourselves between the registry and the docker daemon), we can see the following flow:

CLIENT                                          REGISTRY
(dockerd)                                       (dockerhub... gcr...)

-> GET /v2/
                                                  <- OK


-> GET /v2/file/manifests/latest
                                                  <- manifest


-> GET /v2/file/blobs/sha256:f4f15...
                                                  <- blob content


-> GET /v2/file/blobs/sha256:ffb7f...
                                                  <- blob content

What’s going on there?


checking the version

The first thing docker tries to check is what’s the version the registry understands:

GET /v2/ HTTP/1.1
Host: localhost:5000
User-Agent: docker/18.09.2 ...
Accept-Encoding: gzip
Connection: close

HTTP/1.1 200 OK
Content-Type: application/json; charset=utf-8
Docker-Distribution-Api-Version: registry/2.0   <-------
Connection: close

{}

There’s not much to focus on here aside from:

With the client having validated that it’s interacting with a V2 registry, it can then move forward asking for what it really cares about - the container image.


retrieving the image manifest

The manifest is a JSON file that acts as the provider of:

  1. configuration that describes metadata about that image, and
  2. pointers to where the layers that can be composed to form the filesystem

  MANIFEST ---+----> what this container image is about              (metadata)
              |
              +----> where you can go to get the configuration
              |      for the container that you'll create      (runtime config)
              |      to run with the result of this image
              |
              *----> where you can go get the layers to mount on       (layers)
                     top of each other to form the rootfs

Here’s an example of a manifest of a container image that contains a single layer, no extra metadata, and a runtime configuration file:

{
  "schemaVersion": 2,
  "mediaType": "application/vnd.docker.distribution.manifest.v2+json",
  "config": {
    "mediaType": "application/vnd.docker.container.image.v1+json",
    "size": 1192,
    "digest": "sha256:922f19e5e8f8e734b76618a3e1fe4312c9f07f8d5f83b32c7f33dd9ac38decf7"
  },
  "layers": [
    {
      "mediaType": "application/vnd.docker.image.rootfs.diff.tar",
      "size": 2048,
      "digest": "sha256:e4f8be873d750c0f729a43cac89e1c07a347690c3a8464f35c83109b82b0aa09"
    }
  ]
}

To retrieve that manifest for a given image and a given tag, the following endpoint exists:

GET /v2/<name>/manifests/<reference>

In practice, that turns to the following:


               max size: 256; each field adhering to 
               regex [a-z0-9]+(?:[._-][a-z0-9]+)*
           .---.                
           |   |                
           | .-+-----------------> repository name (could be `foo/bar[/...]`
           | | |                                    as well)
           | | |            .----> reference
           | | |            |   
-> GET /v2/file/manifests/latest

Where the reference is either:

This can be seen in practice when looking at what cartorio does with its internal blobstore:

 blobstore
 ├── bucket
 │   └── sha256:cceeccbdfb5
 └── manifests
     └── file
         ├── latest -> blobstore/bucket/sha256:cceeccbdf
         └── sha256:cceeccbdf -> blobstore/bucket/sha256:cceeccbdfb546


retrieving the image’s runtime config

As you might remember, whenever you have a Dockerfile, you’re able to set runtime configurations:

For example:

{ ... "config": {
    ...
    "AttachStderr": false,
    "Tty": false,
    "OpenStdin": false,
    "StdinOnce": false,
    "Env": [
      "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
    ],
    "Cmd": null,
    ...
    "Labels": null
  }, ...  }

All of these configurations live in the runtime configuration file that gets referenced by in the manifest:

MANIFEST
  {..., "config": {
    "mediaType": "application/vnd.docker.container.image.v1+json",
    "size": 1192,
    "digest": "sha256:922f19e5e8f..."
  }}

As the distribution spec assumes that any blob referenced can be fetched from a blobstore in a content-addressable manner, by following that digest and knowing the mediaType, docker can:

such configuration.

The registry API then requires the existence of a repository-scoped blob retrieval endpoint:

GET /v2/<name>/blobs/<digest>

Which, in practice, looks like:


                    digest      := algorithm ":" hex
                    algorithm   := /[A-Fa-f0-9_+.-]+/
                    hex         := /[A-Fa-f0-9]+/

                    digest: serialized hash result consisting of an algorithm 
                    and hex portion (compliant implementations SHOULD use 
                                     sha256)
                   .----------------------.	 
                   |                      |		[cartorio]: assumes only sha256
GET /v2/repo/blobs/sha256:f4f156284cbb2d...
        |
        |
        repository that "owns" the blob
        (useful for access control)


retrieving the layers

At the same time that the configuration is retrieved, the client who’s performing the pulling can also pull the second types of blobs: layers.

Just like the runtime configuration, these are content-addressable blobs than to be retrieved touching the same blobs endpoint:

GET /v2/<name>/blobs/<digest>

The only difference in this case is that the blob media types (communicated in the manifest beforehand) are going to be different - instead of an container runtime config json, layer differences.