Build Containers the Hard Way (WIP)

Like Kubernetes the Hard Way, but for building containers.

This guide is geared towards anyone interested in learning the low-level details regarding how to build containers. This guide is not intended for those who wish to learn to build container images with high-level tools like Docker.

At the end of this guide, you should understand the internals of a container image, how to construct a container image, and how to push a container image to a Docker container registry piece-by-piece.

What is a container image?

A container is a way of executing processes with isolation provided by 3 Linux technologies - chroot, namespaces, and cgroups.

chroot changes the file system root (/) that a process can see. This allows a process to use any directory as if it were a file system root instead of the actual file system root.

namespaces group resources together (like network and process IDs) so that only processes within a namespace can see the resources of that namespace.

cgroups set CPU and memory limits for processes.

The combination of these allows processes to run in isolation from other processes both on the file system level and on the resource utilization level. Processes inside a container do not see other processes in other containers. Rather, they only see their own view of the system that is not cluttered other processes.

A container image a way to package up an application so that it can run as a container. This package includes the application and any run-time dependencies and is simply a directory of files along with metadata about how to run the container.

A container image is simply a directory of files along with metadata about how to run the container.

Containers from Scratch is a good article explaining how to run your own containers using simple Linux commands.

Build a container image with Docker

Docker is the most popular tool for working with container images. It has built-in support for building and running containers. Docker defines its own scripting language for defining how to build a container image.

For example, this Dockerfile builds a simple Docker image that serves any static files in the current working directory as webpages:

Dockerfile
FROM python
COPY . /public
WORKDIR /public
ENTRYPOINT ["python3", "-m", "http.server"]

To build the Docker image, save the Dockerfile in the current working directory and tell Docker build it:

$ docker build .
Sending build context to Docker daemon  3.072kB
Step 1/4 : FROM python
 ---> 338b34a7555c
Step 2/4 : COPY . /public
 ---> edcd805ec657
Step 3/4 : WORKDIR /public
 ---> Running in cd5f924a79fe
Removing intermediate container cd5f924a79fe
 ---> bba7b6ca42fc
Step 4/4 : ENTRYPOINT ["python", "-m", "http.server"]
 ---> Running in 4198534c1c9c
Removing intermediate container 4198534c1c9c
 ---> 5550043a7340
Successfully built 5550043a7340

The logs of the docker build command clearly show how Docker executes the build based on our Dockerfile.

The docker CLI is the client that you use to send commands and data to the Docker daemon. The Docker daemon stores and runs containers. First, the docker build . command tells Docker to use the local current working directory (.) as the Docker context. The Docker context is sent to the Docker daemon.

The commands in the Dockerfile are then run in the given order. The Docker daemon runs a container to execute each command and generates a new container image at the end of each step.

The FROM python step tells Docker to build the new container starting with the python container image from Docker Hub as the base.

COPY . /public copies all the files in the Docker context into the /public directory on our container image. This includes all the files in our local current working directory that was sent over to the daemon as the Docker context.

WORKDIR /public sets the working directory for the container image we are building, so that when the container image is run, the current working directory for the process is /public.

Note that the container that was created by the WORKDIR command is removed. Steps that do not change the container file system only modify the container configuration and are removed. The container configuration is metadata that describes how to run the container entrypoint process.

ENTRYPOINT ... sets the command to run when the container is run, which, in this case, runs an HTTP server that serves all the files in /public. So, for example, if we had HTML files in our local current working directory, those would be served by this container.

The final image built has ID 5550043a7340, which can be used in subsequent docker commands.

This container can then be run with:

$ docker run -p 8000:8000 5550043a7340

This runs the container and forwards the port to localhost:8000. If you had, say, an index.html in your local current working directory, going to localhost:8000 would serve the contents of index.html.

Another common Dockerfile instruction is RUN <command>, which executes the <command> using whichever shell is present in the container, creating a new container following the result of executing that command.

Advanced Docker Builds

Docker has its own caching mechanism that helps Dockerfile-based builds run faster. Essentially, Docker caches the container image generated after each Dockerfile step. If it determines that a step has not changed, it will use the container image cached for that step rather than run that step again. However, if a step changes, it will invalidate the cache for that step and all steps afterwards. The way the caching mechanism works in Docker merits some tips for writing a more efficient Dockerfile.

Frequently changing steps go last

Dockerfile builds can take quite a while, especially with heavy steps that download dependencies or compile your application.

Docker caches the result of each instruction in a Dockerfile, but once a step's cached result is invalidated, the results of all subsequent steps are invalidated as well.

Since any step that changes invalidates steps afterwards, better Dockerfiles tend to place more frequently-changing steps last. Heavy steps that do not change much tend to be front-loaded so that subsequent builds can just use the cached result.

Keep the number of steps minimal

Since each step generates a separate cached container image, steps should be kept minimal to reduce the number of layers (each cached step has its own overhead). For example, a group of RUNs can usually be combined:

Bad:

RUN mkdir mydirectory
RUN touch mydirectory/myexecutable
RUN chmod +x mydirectory/myexecutable

Better:

RUN mkdir mydirectory && \
    touch mydirectory/myexecutable && \
    chmod +x mydirectory/myexecutable

However, frequently-changing steps should not be combined with infrequently-changing heavy steps - rather, an efficient Dockerfile should try to separate these as much as possible to avoid running infrequently-changing steps.

One important caveat to note is how Docker caches RUN instructions. Even though RUN commands may produce different results each time, Docker only invalidates the cache for a RUN instruction if the command itself changes. This may produce unexpected results if you intended a RUN instruction to not use a cached version.

Use multi-stage builds for more flexibility than just a linear Dockerfile build

Single-stage Dockerfile builds (builds with a single FROM instruction) builds a single image starting with the FROM image as the base. These builds run in a purely linear format from start to finish and the results of all intermediate steps become part of the resulting image. However, since a container image should only contain what is necessary to run your application, much of the files from intermediate steps are junk that is not needed at run-time.

For example, for a Java (Maven) project, you may need the JDK to build the JAR, but only the JRE to run the JAR. The multi-stage Dockerfile build would look something like:

FROM maven AS builder
COPY . .
RUN mvn package

FROM openjdk:8-jre-alpine
COPY --from=builder target/my-app-*.jar app.jar
ENTRYPOINT ["java", "-jar", "app.jar"]
  1. This multi-stage build starts with the Docker Hub maven image as the base for an intermediate image called builder.

  2. It then copies all the project files into the builder image.

  3. mvn package builds the fatjar for the project.

  4. The main stage starts with the Docker Hub openjdk image that just has the JRE.

  5. It copies just the fatjar from the builder image.

  6. And sets the container to run that fatjar when executed.

The resulting container contains only what is necessary to run the application - the application JAR and the Java run-time.

You can also leverage multi-stage builds to take advantage of better caching, since you're not limited by a purely linear cache invalidation.

Container image format

Container images can be built in a few formats. Different runtimes may support different formats. The most common formats are the Docker image format and Open Container Initiative (OCI) format.

Some common runtimes include:

There are also many tools for building images besides Docker:

  • Jib builds Docker and OCI images in Java

  • Kaniko builds images in Kubernetes using a Dockerfile

  • rules_docker provides Bazel rules for building images

  • BuildKit is the underlying engine used by Docker to build images

  • img provides a standalone frontend for BuildKit

  • buildah builds OCI images

Layers

As mentioned, a container image is simply a directory of files (along with metadata). In the image format, this directory of files is split up into layers that are compatible with union mounting.

Layers are independent, but are composed together with a defined order in the image format. The order of the layers in the image matters for files that exist in multiple layers. The contents of a file that appears in a later layer replaces the contents of the same file in a prior layer. A file can also be deleted in a later layer by prepending its filename with .wh..

Each layer is archived into a single tarball. In the Docker image format, this tarball must be compressed with GZIP compression.

Note that when a container image is run, Docker shares common layers between containers. Docker also creates an overlay layer for file system writes - any writes to existing files would be a copy-on-write to this overlay layer. When the container is removed, this overlay layer is deleted.

TODO: Show some bash commands that can generate a layer

Digests

In the image format, BLOBs (such as the layer archives) are identified by their digests. These digests are generated by hashing the BLOB contents. The digest for a layer is the hash of the compressed tarball archive (compressed with GZIP). The hashing algorithm to use for Docker image format is SHA-256, but the OCI format supports SHA-512 as well. The benefit of identifying BLOBs is that they become content-addressable. Uniqueness is provided by the collision-resistance of SHA-256.

Digests are usually written with the hashing algorithm as the prefix, like sha256:0b56ce23cf77e0a18c94b4ca8fe5fa58f4cca119f61ba4021cde2c9db56faaa4.

Container configuration

The metadata for a container image is contained its container configuration. This container configuration contains mainly information about how the container should be run. This includes fields such as for architecture and OS (GOARCH/GOOS), exposed ports, volume mounts, and environment variables.

This is an example of a container configuration:

{
  "architecture": "amd64",
  "os": "linux",
  "config": {
    "Env": ["PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"],
    "Entrypoint": ["/usr/bin/java"],
    "Cmd": ["arg1", "arg2"],
    "ExposedPorts": { "6000/tcp":{}, "8000/tcp":{}, "9000/tcp":{} },
    "Labels": { "com.example.label": "value" },
    "WorkingDir": "/home/user/workspace",
    "User": "me"
  },
  "history": ...,
  "rootfs": {
    "diff_ids": [
      "sha256:2aebd096e0e237b447781353379722157e6c2d434b9ec5a0d63f2a6f07cf90c2",
      "sha256:5f70bf18a086007016e948b04aed3b82103a36bea41755b6cddfaf10ace3c6ef",
    ],
    "type": "layers"
  }
}

The architecture and os fields define the platform this container is built to run on. The config field defines the runtime configuration for the container, including any environment variables (Env), the entrypoint (Entrypoint), arguments to append to the entrypoint (Cmd), ports to expose (ExposedPorts), arbitrary annotations (Labels), the directory to use as the current working directory (WorkingDir), and the user to run the container process as (User).

The most important configuration is the entrypoint, which is the command that is run when the container is run. This essentially defines the entrypoint process the container is isolating.

Note that all run-time configuration in the container configuration exist as defaults. When actually running the container either with Docker or with Kubernetes, any/all of these run-time configurations can be overridden.

TODO: Might want to talk about reasonings for burning runtime config into an image vs. defining in a runtime spec

TODO: Show example container configuration and running with Docker/Kubernetes and effective runtime configuration

There are two other parts of the container configuration: the rootfs object and the history array. Both contain metadata about the layers that make up the container file system.

The rootfs contains a list of diff IDs. A diff ID for a layer is the digest (usually SHA256 hash) of the uncompressed tarball archive containing the files for that layer. Note that this is different from a layer's descriptor digest, which is a hash of the compressed archive. The rootfs defines a list of these diff IDs in the order in which the layers belong in the container overlay file system, from first to last. Note that these layers must match those defined in the manifest.

Manifest

A container image manifest describes the components that make up a container image. Manifests come in multiple forms. These forms include the registry manifest formats such as Docker Image Manifest V 2, Schema 2 and OCI Image Manifest, as well as manifest formats that the Docker toolchain uses to save and load images stored locally.

The main information in a manifest are the components that make up a container image. These components include the layers and the container configuration.

docker load format

docker load can load a container image stored as a tarball archive into the Docker local repository store. This tarball archive includes the compressed tarball archives for all the layers, the container configuration JSON file, and the manifest JSON file (must be named manifest.json). Here’s an example of the manifest JSON file:

[
  {
    "Config":"config.json",
    "RepoTags":["myimage"],
    "Layers": [
      "layer1.tar.gz",
      "layer2.tar.gz",
      "layer3.tar.gz"
    ]
  }
]

The reference for this format can be found in the Docker codebase: https://github.com/moby/moby/blob/master/image/tarexport/load.go

When Docker loads this image tarball, Docker finds this manifest.json and reads it. The manifest tells Docker to load the container configuration from the config.json file and load the layers from layer1.tar.gz, layer2.tar.gz, and layer3.tar.gz, in that order. Note that this order must match the order of the layers defined in the container configuration under rootfs. The RepoTags here tells Docker to "tag" the image with the name myimage.

Note that "tag" is a confusing term here since it refers to the full reference for the image, whereas "tag" in an actual image reference refers to a label for an image stored under a repository.

docker save format

docker save can also save images in a legacy Docker tarball archive format that docker load also supports. In this legacy format, each layer would be stored in its own directory named with its SHA256 hash (digest of compressed layer tarball archive). Each of these directories contains a json file with a legacy container configuration (we won’t go into the details of this), a VERSION file, and a layer.tar that is the uncompressed tarball archive of the layer contents.

The reference for this format can be found in the Docker codebase: https://github.com/moby/moby/blob/master/image/spec/v1.2.md

Registry format - Docker Image Manifest V 2, Schema 2

Registry image manifests define the components that make up an image on a container registry (see section on container registries). The more common manifest format we’ll be working with is the Docker Image Manifest V2, Schema 2 (more simply, V2.2). There is also a V2, Schema 1 format that is commonly used but more complicated than V2.2 due to backwards-compatibility reasons against V1.

The V2.2 manifest format is a JSON blob with the following top-level fields:

schemaVersion - 2 in this case

mediaType - application/vnd.docker.distribution.manifest.v2+json

config - descriptor of container configuration blob

layers - list of descriptors of layer blobs, in the same order as the rootfs of the container configuration

Blob descriptors are JSON objects containing 3 fields:

mediaType - application/vnd.docker.container.image.v1+json for a container configuration or application/vnd.docker.image.rootfs.diff.tar.gzip for a layer

size - the size of the blob, in bytes

digest - the digest of the content

Here is an example of a V2.2 manifest format (for the Docker Hub busybox image):

{
  "schemaVersion": 2,
  "mediaType": "application/vnd.docker.distribution.manifest.v2+json",
  "config": {
    "mediaType": "application/vnd.docker.container.image.v1+json",
    "size": 1497,
    "digest": "sha256:3a093384ac306cbac30b67f1585e12b30ab1a899374dabc3170b9bca246f1444"
  },
  "layers": [
    {
      "mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
      "size": 755724,
      "digest": "sha256:57c14dd66db0390dbf6da578421c077f6de8e88edd0815b4caa94607ba5f4c09"
    }
  ]
}

This manifest states that the busybox image is composed of a container configuration stored as a blob with digest sha256:3a093384… and a single layer blob with digest sha256:57c14dd6…. Note that manifests simply contain references. The actual blob content is stored elsewhere and would need to be fetched and provided to a container run-time when running the image as a container.

It is important to make sure that the digest used in the descriptor for a layer is the digest of the compressed archive, whereas the digest used for the diff ID in the rootfs of the container configuration is the digest of the same archive but uncompressed. Otherwise, container registries may give you a cryptic MANIFEST_INVALID error.

The reference for this format can be found in the Docker documentation: https://docs.docker.com/registry/spec/manifest-v2-2/

Registry format - OCI Image Manifest

The OCI image manifest is also a registry image manifest that defines components that make up an image. The format is essentially the same as the Docker V2.2 format, with a few differences.

mediaType - must be set to application/vnd.oci.image.manifest.v1+json

config.mediaType - must be set to application/vnd.oci.image.config.v1+json

Each object in layers must have mediaType be either application/vnd.oci.image.layer.v1.tar+gzip or application/vnd.oci.image.layer.v1.tar.

OCI also allows for layer blobs to be not compressed with GZIP with the application/vnd.oci.image.layer.v1.tar mediaType.

The reference for this format can be found in the OCI codebase: https://github.com/opencontainers/image-spec/blob/master/manifest.md

Manifest List/Image Index

A manifest list (or image index in OCI terms) is a way to specify multiple manifests for a single image. Each manifest is associated with a specific platform (operating system and architecture) so that a container run-time can choose the appropriate manifest to use for the platform it is running on. Manifest lists are fairly uncommon and we will not go into details about them. See official documentation for more information about a manifest list or image index.

Let’s explore a Docker image with docker

TODO

Container image registry

In order for a container orchestration system like Kubernetes to run container images, it needs to pull the container images from some container image repository. This is where container image registries come in. Container image registries store container images to be served via specific addresses called image references. These registries are usually BLOB stores that store the container image layers in separate BLOBs. They also store manifests that describe which blobs make up a specific image. The most common container image registry is the Docker Registry V2, which is an API specification for a registry server. Some examples of hosted Docker registries include Docker Hub (the default image registry for Docker), Google Container Registry (GCR), AWS Elastic Container Registry (ECR), Microsoft Azure Container Registry (ACR), Harbor, JFrog Artifactory, Quay, and Sonatype Nexus.

We will go over the details of Docker Registry API V2 (we’ll just refer to it as the Registry API), since most container registries implement that specification. For the full specification details, read the Docker Registry HTTP API V2 spec.

How a registry stores an image

Although the specific implementation details are undefined and differ between each implementation of the Registry API, conceptually, the registry can be thought of as a blob store that can have blobs pushed to and pulled from. The Registry API works with the Docker/OCI Image Format, so the blobs would be the layers and the container configurations. The manifests are not referred to as blobs, but rather are the artifacts that actually represent an image (by listing its component blobs).

The diagram shows the topology of a registry. A registry contains multiple repositories containing images. Each repository contains a set of blobs. An “image” is actually a manifest that refers to a sequence of blobs that make up the image. Tags are labels used to name manifests. A tag can only point to a single manifest but multiple tags can point to the same manifest.

Image reference

An image is identified by an image reference. This image reference provides a human-readable way to address images and is composed of a few components:

  • Registry - The URL to reach the registry server. By default, this registry would be Docker Hub (registry.hub.docker.com).

  • Repository - Also known as the namespace, repositories help organize images into their own “directories” on the registry. Authentication applies on a per-repository basis. When the registry is Docker Hub, the default repository prefix for a single-level repository is library/ (for example, busybox refers to library/busybox).

  • Tag - Each image within a repository is addressable by the digest of its manifest. However, digests can be difficult to manage and is not human-friendly. Tags can be set to point to specific digests to more conveniently identify specific images. These tags can also be reassigned to different digests as well. A specific image can also have multiple tags pointing to it. Each new manifest must go under a specific tag and the default tag is latest.

For example, consider the following image reference: openjdk:8-jre-alpine

  • Since there is no registry URL specified, the registry is by default Docker Hub. The repository component is specified as openjdk. Since the repository is only a single level and the registry is Docker Hub, the repository is actually resolved as library/openjdk. The tag component is specified after a colon. In this example, the tag is 8-jre-alpine. Although this tag is arbitrarily-defined, the maintainers of this repository chose to convey some information about the image through this tag. Here, this tag says that the specific image contains OpenJDK 8 with just the JRE (Java Runtime Environment) and Alpine (a tiny Linux distribution). Some other tags in the repository refer to the same image but contain more specific information, like 8u191-jre-alpine3.9. Other tags may be less specific, like jre, which, at the time of writing, refers to an OpenJDK 11 image with the JRE and Debian Stretch. This tag will most likely move to refer to newer OpenJDK versions as they become GA. The latest tag also refers to the latest stable, default image the maintainers believes users will find useful in most cases and will move the tag as new versions become stable. There are some general tips and best practices for managing tags for a repository.

TODO: Maybe explain some tips?

Consider another image reference: gcr.io/my-gcp-project/my-awesome-app/message-service:v1.2.3

  • The registry URL here is gcr.io, which is the server URL for Google Container Registry. The repository is my-gcp-project/my-awesome-app/message-service, which in GCR, means that the repository is on the my-gcp-project Google Cloud Platform project and under a my-awesome-app/message-service Google Cloud Storage subdirectory. The tag is v1.2.3, which is used to identify the version of the message-service this container image contains.

An image can also be referred to by its specific image digest, such as gcr.io/distroless/java@sha256:0430beea631408faa266e5d98d9c5dcc3c5f02c3ebc27a65d72bd281e3576097.

There are also specific acceptable patterns for each component. The full regex can be found in the Docker distribution code. For a brief summary:

  • Registry - can be a standard DNS hostname (no protocol) with an optional port number

  • Repository - slash-separated components that contain lowercase letters and digits separated by period, one or two underscores, or one or more dashes

  • Tag - any of letters, digits, underscores, periods, and dashes for a maximum of 128 characters

Note that "tag" is a confusing name in Docker terminology since when using the local docker CLI tool, "tag" also refers to a label given to an image on the Docker daemon. For example, you might build an image called myimage with:

$ docker build -t myimage .
...
Successfully built 4e905a76595b
Successfully tagged myimage:latest

Docker uses the current directory as the context and builds an image with an image ID 4e905a76595b. This image ID is a shortened SHA256 digest of the built container configuration. Docker then "tags" the 4e905a76595b with a label myimage, which can be used to refer to the same image with future docker commands rather than using 4e905a76595b all the time. However, when you want to build and push an image to a registry, you would "tag" the image with the full image reference:

$ docker build -t gcr.io/my-gcp-project/myimage:v1 .
...
Successfully built 4e905a76595b
Successfully tagged gcr.io/my-gcp-project/myimage:v1
$ docker push gcr.io/my-gcp-project/myimage:v1

Here, Docker "tags" the image with gcr.io/my-gcp-project/myimage:v1 to be used in later docker commands, but gcr.io/my-gcp-project/myimage:v1 is actually an image reference with its tag as v1.

Pulling an image

Pulling an image involves two steps:

  1. Pull the manifest.

  2. Pull the blobs in the manifest.

Pull the manifest

To pull the manifest, send a request to:

GET /v2/<repository>/manifests/<tag or digest>

For example, to get the manifest for openjdk, the endpoint would be https://registry.hub.docker.com/v2/library/openjdk/manifests/latest.

Headers to send include Authorization (explained in the Token Authentication section) and Accept. For example, if you want to only accept and parse manifests for Docker Image Format V2 Schema 2, you would want to set Accept: application/vnd.docker.distribution.manifest.v2+json.

Manifests can also come in list format for some images. These would have a media type like application/vnd.docker.distribution.manifest.list.v2+json. These manifests lists contain specific manifests for different architectures and operating systems. For example, if you are looking specifically for a manifest for the amd64/Linux platform, you would parse the manifest list for the corresponding manifest for that platform and then pull that manifest using its specific digest. See an example of a manifest list.

Pull the blobs

After the manifest is pulled, the specific blob digests can be extracted for the container configuration and the layers. Each blob can be pulled in parallel via:

GET /v2/<repository>/blobs/<digest>

You might want to verify that the size of the blob pulled is what you expected based on the size specified in the manifest.

Pushing an image

Pushing an image is like pulling an image except in reverse:

  1. Push the blobs.

  2. Push the manifest.

Push the blobs

Blobs can be pushed in parallel and pushing each blob is a 3-step process:

  1. Initialize upload.

  2. Send content.

  3. Commit with digest.

Initialize the upload by sending a request to:

POST /v2/<repository>/blobs/uploads/

A successful 202 Accepted response should contain a Location header with the URL to send the blob content to. There are a few ways to send the blob content, but the recommended way is to send the content in a single request. The content can also be sent in chunks if you wish to implement resumable uploads. Send the blob content to the Location via:

PATCH <Location>

<blob binary data>

The success response should be 202 Accepted. Once a blob is uploaded, the blob can be committed with its digest:

PUT <Location>?digest=<digest>

The success response should be 201 Created.

Avoid unnecessary pushes

Check if blob already exists

To avoid pushing a blob that already exists on a registry, if the blob’s digest is known beforehand, you can check the existence of the blob with:

HEAD /v2/<repository>/blobs/<digest>

If the blob exists, the response will be 200 OK.

Cross-repository blob mount

If the blob exists in another repository in the same registry, you can also try directly mounting the blob into the target repository:

POST /v2/<repository>/blobs/uploads/?mount=<digest>&from=<source repository>

If the blob mount succeeds, the response will be 201 Created. If the blob mount fails (blob doesn’t exist or cross-repository blob mount not supported), the response will be a normal blob upload initialization response with a Location to send the blob content to.

An important aspect to note is that the Authorization to send with this request must have permissions to pull from the source repository in addition to permissions to push to the target repository. See the Token Authentication section below for details.

Push the manifest

Once all the blobs are committed, the manifest can be pushed via:

PUT /v2/<repository>/manifests/<tag>

Make sure to set the Content-Type to the correct manifest media type. For Docker V2.2, the media type would be application/vnd.docker.distribution.manifest.v2+json and for OCI, the media type would be application/vnd.oci.image.manifest.v1+json.

Authenticating requests

Most registry requests would require a sufficient Authorization header to be authenticated. Otherwise, a 401 Unauthorized or 403 Forbidden response would be received.

The standard flow for any request should be:

  1. Try the request.

  2. If unauthorized, obtain a token.

  3. Retry the request with the token.

For example, an authorized response may include a WWW-Authenticate header like:

Www-Authenticate: Bearer realm="https://auth.docker.io/token",service="registry.docker.io",scope="repository:coollog/myimage:pull,push"

To obtain a token, send the following request along with an Authorization header containing the credentials for access to the repository:

GET https://auth.docker.io/token?service=registry.docker.io&scope=repository:coollog/myimage:pull,push

For example, if the Docker Hub password for the coollog account was donthackme123, the Authorization header would be Authorization: Basic Y29vbGxvZzpkb250aGFja21lMTIz, where Y29vbGxvZzpkb250aGFja21lMTIz is the Base64 encoding of coollog:donthackme123.

The response should be a JSON containing either a token or access_token (OAuth2) that is the authenticated Bearer token. The JSON also may also contain an expires_in and issued_at field for knowing when the token expires. These tokens should be cached and applied to any subsequent requests (until expired) to skip steps 1 and 2.

Note that multiple scopes can be included. For example, for a cross-repository blob mount with otheruser/baseimage as the source repository, the request may be:

GET https://auth.docker.io/token?service=registry.docker.io&scope=repository:coollog/myimage:pull,push&scope=repository:otheruser/baseimage:pull

This request would need to include Authorization for both scopes though.

Where to find registry credentials

If a user has Docker installed, they most likely have it configured with credentials to use for various repositories they have access to (through docker login). This configuration can be found at the .docker/config.json file located in the user’s home directory. This configuration can store both raw credentials and references to credential helpers. Credential helpers are CLI tools that can be used to fetch credential.

The raw credentials are stored in an auths field that is a map with registry URLs as keys. For example:

{
  "auths": {
    "https://gcr.io": {
      "email": "not@val.id",
      "auth": "<Base64-encoded username:password>"
    }
  }
}

The credential helper references are stored in a credHelpers field that maps from registry URLs to credential helper names. For example:

{
  "credHelpers": {
    "gcr.io": "gcr",
    "us.gcr.io": "gcr",
    "aws_account_id.dkr.ecr.region.amazonaws.com": "ecr-login"
  }
}

A default credential helper can also be specified in a credsStore field. For example:

{
  "credsStore": "osx-keychain"
}

Note that registry URLs can come with or without a protocol and with or without arbitrary suffixes. For example, the following all refer to gcr.io:

gcr.io
https://gcr.io
gcr.io/v2/
https://gcr.io/v2/

Credential helpers are command-line tools that can be used to fetch credentials for a registry. To fetch credentials from a credential helper, call the credential helper get command, passing in the server URL as the input. For example, to fetch credentials from the credential helper named gcr for the gcr.io registry, call:

$ echo gcr.io | docker-credential-gcr get

If the credentials are present, the output should be a JSON that includes a Username and Secret field which can be used as the username and password, respectively, for authenticating a token.

Note that when the Username returned is <token>, the Secret is actually an identity token that should be used as a refresh token in obtaining an OAuth2 access token. For example, Azure's credential helper returns a JSON that looks like:

{
    "ServerURL": "",
    "Username": "<token>",
    "Secret": "<eyJhbGciOi...88Q>"
}

The Secret should be used as the refresh token for an OAuth2 access token request:

POST https://myregistry.azurecr.io/oauth2/token
Content-Type: application/x-www-form-urlencoded

grant_type=refresh_token&service=myregistry.azurecr.io&scope=repository:andy:pull&refresh_token=<eyJhbGciOi...88Q>

Local caching

When building an image by constructing layers on top of a base image, it may be expensive to pull base image layers every time. To resolve this, you should use a local cache to store images.

For example, Jib uses a simple cache mechanism that stores blobs in a specific directory structure. The cache directory contains two subdirectories: blobs/ and selectors/. The blobs/ directory contains files for each blob named by their digest. The selectors/ directory contains selector files whose names are the selectors themselves and contents are the digests of the blobs they point to. Both blobs and selectors should be immutable and unique.

This design supports fetching a blob by digest and by selector. When writes are made atomic, the cache can be used fully concurrently across multiple blobs. Selectors are useful for “finding” a blob by some representation of its contents. For example, in Jib, some blobs are layers that were built from the Java application source code. The selectors for these blobs are a list of files that went into building that layer. The name of a selector is the SHA256 digest of that file list. Uniqueness is provided by associating each file with its last modified time and immutability is provided by the determinism of SHA256. What this allows Jib to do is that when files don’t change, they will always generate the same selector and give back an existing blob so that that blob does not need to be re-generated; but, when files do change, the generated selector is different and a new selector-blob pair will be added to the cache.

An example cache directory layout could look like:

blobs/
  36a2b7401dcddc50a35aeaa81085718b9d5fbce9d607c55a1d79beec2469f9ac
  ...
selectors/
  65de3b72aaf98e4f300ccdf7d64bf9a3b1e23c8c44a1242265f717db1a0877e9
  ...

Here, the 65de3b72… selector file contains the 36a2b740… digest. See the proposal for the original Jib cache design for more details.

Large caches

As caches grow in size, they may need to be pruned. A simple mechanism is to use a least-recently-used method. For caches that grow large, however, resolving a blob or selector may become expensive since directory traversal is at best logarithmic to the number of directory entries. A simple technique to alleviate this would be to shard the directory into a number of bins, and potentially do so recursively within each bin. For example, the bins could be named 00 through ff, with each file being placed into the bin that matches the first two characters in its file name.

Walk through: pulling an image with bash

Let's walk through pulling the busybox image from Docker Hub using some bash commands.

First, we will try to obtain a token to pull the busybox image. We curl the registry API root to get the WWW-Authenticate challenge:

$ curl -v https://registry.hub.docker.com/v2/
...
< HTTP/1.1 401 Unauthorized
...
< Www-Authenticate: Bearer realm="https://auth.docker.io/token",service="registry.docker.io"
...

Okay, we can now respond to this challenge to get an access token for busybox. Since busybox is a public image, we don't need to include any Authorization.

$ curl -v "https://auth.docker.io/token?service=registry.docker.io&scope=repository:library/busybox:pull"
...
< 
{
  "token": "eyJhbGci...",
  "access_token": "eyJhbGci...",
  ...
}

We can now store this token and use it to obtain the manifest for busybox:

$ TOKEN=eyJhbGci...
$ curl https://registry.hub.docker.com/v2/library/busybox/manifests/latest \
      -H "Accept: application/vnd.docker.distribution.manifest.v2+json" \
      -H "Authorization: Bearer $TOKEN"
{
  "schemaVersion": 2,
  "mediaType": "application/vnd.docker.distribution.manifest.v2+json",
  "config": {
    "mediaType": "application/vnd.docker.container.image.v1+json",
    "size": 1497,
    "digest": "sha256:d8233ab899d419c58cf3634c0df54ff5d8acc28f8173f09c21df4a07229e1205"
  },
  "layers": [
    {
      "mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
      "size": 755901,
      "digest": "sha256:697743189b6d255069caf6c455be10c7f8cae8076c6f94d224ae15cd41420e87"
    }
  ]
}

Great, now let's fetch the container configuration:

$ curl https://registry.hub.docker.com/v2/library/busybox/blobs/sha256:d8233ab8... \
       -H "Authorization: Bearer $TOKEN" \
       -L \
       -o config.json
$ cat config.json
{
  "architecture": "amd64",
  "config": ...,
  "container": ...,
  "container_config": ...,
  "created": ...,
  "docker_version": ...,
  "history": ...,
  "os": "linux",
  "rootfs": {
    "type": "layers",
    "diff_ids": [
      "sha256:adab5d09ba79ecf30d3a5af58394b23a447eda7ffffe16c500ddc5ccb4c0222f"
    ]
  }
}

And also the layer archive:

$ curl https://registry.hub.docker.com/v2/library/busybox/blobs/sha256:69774318... \
      -H "Authorization: Bearer $TOKEN" \
      -L \
      -o layer.tar.gz

Now, let's write the manifest.json for docker load:

$ echo '[
  {
    "Config":"config.json",
    "RepoTags":["busybox:latest"],
    "Layers": [
      "layer.tar.gz"
    ]
  }
]' > manifest.json

We can now tar up the entire image and load it into Docker:

$ tar -cvf busybox.tar *
$ docker load < busybox.tar
adab5d09ba79: Loading layer [==================================================>]  755.9kB/755.9kB
Loaded image: busybox:latest

We can now run the busybox image we just pulled:

$ docker run -it busybox
/ # ls
bin   dev   etc   home  proc  root  sys   tmp   usr   var

Walk through: building and pushing an image with bash

Now, let's try building an image with bash. We'll build a simple image that has a shell script that prints "Hello World" on top of the busybox image.

First, let's write the "Hello World" shell script and archive it into its own layer tarball:

$ echo 'echo "Hello World"' > hello.sh && chmod +x hello.sh
$ tar -czvf app-layer.tar.gz hello.sh

We can get the SHA256 digest for this layer, its diff ID, and size in bytes:

$ sha256sum < app-layer.tar.gz 
c6c6bfc03cd5df3c175553d2f6c178fb57c9430841a5cb18b6bef9a224849e3c  -
$ gunzip < app-layer.tar.gz | sha256sum
9c20a6aa71f1cd3cc2f794e235cc505403a251b6ebce7f90418411e2142fe32f  -
$ stat -c%s app-layer.tar.gz
162

Next, we'll create the container configuration and get its digest/size:

$ echo '{
  "architecture": "amd64",
  "os": "linux",
  "config": {
    "Entrypoint": ["sh", "hello.sh"]
  },
  "rootfs": {
    "type": "layers",
    "diff_ids": [
      "sha256:adab5d09ba79ecf30d3a5af58394b23a447eda7ffffe16c500ddc5ccb4c0222f",
      "sha256:9c20a6aa71f1cd3cc2f794e235cc505403a251b6ebce7f90418411e2142fe32f"
    ]
  }
}' > config.json
$ sha256sum < config.json
586acba684019cc55f34a06f0f86b1864033f4ec4a4757366dcf4a420543ed07  -
$ stat -c%s config.json
329

Now we can construct the manifest:

$ echo '{
  "schemaVersion": 2,
  "mediaType": "application/vnd.docker.distribution.manifest.v2+json",
  "config": {
    "mediaType": "application/vnd.docker.container.image.v1+json",
    "size": 329,
    "digest": "sha256:586acba684019cc55f34a06f0f86b1864033f4ec4a4757366dcf4a420543ed07"
  },
  "layers": [
    {
      "mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
      "size": 755901,
      "digest": "sha256:697743189b6d255069caf6c455be10c7f8cae8076c6f94d224ae15cd41420e87"
    },
    {
      "mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
      "size": 162,
      "digest": "sha256:c6c6bfc03cd5df3c175553d2f6c178fb57c9430841a5cb18b6bef9a224849e3c"
    }
  ]
}' > manifest.json

Let's say we want to push to our own Docker Hub repository. If you don't have a Docker Hub account, simply register one. Let's push our image to <username>/hello. First, we need to push our layers. The first layer comes from busybox, so we can actually just mount that into our new image:

$ USERNAME=<username>
$ curl "https://auth.docker.io/token?service=registry.docker.io&scope=repository:$USERNAME/hello:pull,push&scope=repository:library/busybox:pull" \
      -u "$USERNAME" \
      -v
{
  "token": "eyJhbGci...",
  ...
}
$ TOKEN=eyJhbGci...
$ curl "https://registry.hub.docker.com/v2/$USERNAME/hello/blobs/uploads/?mount=sha256:697743189b6d255069caf6c455be10c7f8cae8076c6f94d224ae15cd41420e87&from=library/busybox" \
      -H "Authorization: Bearer $TOKEN" \
      -d '' \
      -v
...
< HTTP/1.1 201 Created
...

Cool, we successfully mounted the busybox layer into our <username>/hello repository.

The token most likely has a 5-minute expiration, so you may need to re-authenticate if it expires or your requests may fail with 401 Unauthorized.

Now, let's push our hello.sh layer:

$ curl https://registry.hub.docker.com/v2/$USERNAME/hello/blobs/uploads/ \
      -H "Authorization: Bearer $TOKEN" \
      -d '' \
      -v
...
< HTTP/1.1 202 Accepted
...
< Location: https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/d3221a35-5c4e-45c5-9a1a-8f274bfbe077?_state=_7kmmic...
...

We can send the layer contents to the Location received:

$ LOCATION=https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/d3221a35-5c4e-45c5-9a1a-8f274bfbe077?_state=_7kmmic...
$ curl $LOCATION \
      -X PATCH \
      --data-binary @app-layer.tar.gz \
      -H "Content-Type: application/octet-stream" \
      -H "Authorization: Bearer $TOKEN" \
      -v
...
< HTTP/1.1 202 Accepted
...
< Location: https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/51e06c34-74ee-4c65-bc59-b8fbfa79f916?_state=iQtNKd...
...

We then commit our BLOB to the new Location received:

$ LOCATION=https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/51e06c34-74ee-4c65-bc59-b8fbfa79f916?_state=iQtNKd...
$ curl "$LOCATION&digest=sha256:c6c6bfc03cd5df3c175553d2f6c178fb57c9430841a5cb18b6bef9a224849e3c" \
      -X PUT \
      -H "Authorization: Bearer $TOKEN" \
      -v
...
< HTTP/1.1 201 Created
...

Great, now that we have our layer uploaded, we can do the same for our container configuration BLOB:

$ curl https://registry.hub.docker.com/v2/$USERNAME/hello/blobs/uploads/ \
      -H "Authorization: Bearer $TOKEN" \
      -d '' \
      -v
...
< Location: https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/c2be3c4b-d773-43d8-b1bd-b94eab1e20a0?_state=ZYOlPQD...
...
$ LOCATION=https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/c2be3c4b-d773-43d8-b1bd-b94eab1e20a0?_state=ZYOlPQD...
$ curl $LOCATION \
      -X PATCH \
      --data-binary @config.json \
      -H "Content-Type: application/octet-stream" \
      -H "Authorization: Bearer $TOKEN" \
      -v
...
< Location: https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/c2be3c4b-d773-43d8-b1bd-b94eab1e20a0?_state=xsZ-rOO...
...
$ LOCATION=https://registry.hub.docker.com/v2/<username>/hello/blobs/uploads/c2be3c4b-d773-43d8-b1bd-b94eab1e20a0?_state=xsZ-rOO...
$ curl "$LOCATION&digest=sha256:586acba684019cc55f34a06f0f86b1864033f4ec4a4757366dcf4a420543ed07" \
      -X PUT \
      -H "Authorization: Bearer $TOKEN" \
      -v
...
< HTTP/1.1 201 Created
...

Now, we can push our manifest:

$ curl https://registry.hub.docker.com/v2/$USERNAME/hello/manifests/latest \
      -X PUT \
      --data-binary @manifest.json \
      -H "Content-Type: application/vnd.docker.distribution.manifest.v2+json" \
      -H "Authorization: Bearer $TOKEN" \
      -v
...
< HTTP/1.1 201 Created
...

Congrats! You've just built and pushed your own container image to a container registry! Now, you can try pulling and running your image:

$ docker pull $USERNAME/hello
$ docker run $USERNAME/hello
Hello World

Making an efficient container builder

TODO

  • Separate into fine layers

  • Speed up iterative build cycle - separate out parts that change more often into another layer

  • Minimize change

  • stream everything through sinks/sources

Last updated