/official-images

Primary source of truth for the Docker "Official Images" program

Primary LanguageShellApache License 2.0Apache-2.0

Docker Official Images

Table of Contents

  1. Docker Official Images
    1. Table of Contents
    2. What are "Official Images"?
    3. Architectures other than amd64?
    4. More FAQs?
    5. Contributing to the standard library
      1. Review Guidelines
        1. Maintainership
        2. Repeatability
        3. Consistency
        4. Clarity
        5. init
        6. Cacheability
        7. Security
          1. Image Build
          2. Runtime Configuration
          3. Security Releases
        8. Multiple Architectures
      2. Commitment
    6. Library definition files
      1. Filenames
      2. Tags and aliases
      3. Instruction format
      4. Creating a new repository
      5. Adding a new tag in an existing repository (that you're the maintainer of)
      6. Change to a tag in an existing repository (that you're the maintainer of)
    7. Bashbrew

What are "Official Images"?

The Docker Official Images are curated images hosted on Docker Hub. The main tenets are:

See Docker's documentation for a good high-level overview of the program.

In essence we strive to heed upstream's recommendations on how they intend for their software to be consumed. Many images are maintained in collaboration with the relevant upstream project if not maintained directly by them. Additionally we aim to exemplify the best practices for Dockerfiles to serve as a reference when making or deriving your own images from them.

(If you are a representative of an upstream for which there exists an image and you would like to get involved, please see the Maintainership section below!)

Architectures other than amd64?

Some images have been ported for other architectures, and many of these are officially supported (to various degrees).

As of 2017-09-12, these other architectures are included under the non-prefixed images via "manifest lists" (also known as "indexes" in the OCI image specification), such that, for example, docker run hello-world should run as-is on all supported platforms.

If you're curious about how these are built, head over to https://doi-janky.infosiftr.net/job/multiarch/ to see the build scaffolding.

See the multi-arch section below for recommendations in adding more architectures to an official image.

More FAQs?

Yes! We have a dedicated FAQ repository where we try to collect other common questions (both about the program and about our practices).

Contributing to the standard library

Thank you for your interest in the Docker official images project! We strive to make these instructions as simple and straightforward as possible, but if you find yourself lost, don't hesitate to seek us out on Freenode IRC in channel #docker-library or by creating a GitHub issue here.

Be sure to familiarize yourself with Official Repositories on Docker Hub and the Best practices for writing Dockerfiles in the Docker documentation. These will be the foundation of the review process performed by the official images maintainers. If you'd like the review process to go more smoothly, please ensure that your Dockerfiles adhere to all the points mentioned there, as well as below, before submitting a pull request.

Also, the Hub descriptions for these images are currently stored separately in the docker-library/docs repository, whose README.md file explains more about how it's structured and how to contribute to it. Please be prepared to submit a PR there as well, pending acceptance of your image here.

Review Guidelines

Because the official images are intended to be learning tools for those new to Docker as well as the base images for advanced users to build their production releases, we review each proposed Dockerfile to ensure that it meets a minimum standard for quality and maintainability. While some of that standard is hard to define (due to subjectivity), as much as possible is defined here, while also adhering to the "Best Practices" where appropriate.

A checklist which may be used by the maintainers during review can be found in NEW-IMAGE-CHECKLIST.md.

Maintainership

Version bumps and security fixes should be attended to in a timely manner.

If you do not represent upstream and upstream becomes interested in maintaining the image, steps should be taken to ensure a smooth transition of image maintainership over to upstream.

For upstreams interested in taking over maintainership of an existing repository, the first step is to get involved in the existing repository. Making comments on issues, proposing changes, and making yourself known within the "image community" (even if that "community" is just the current maintainer) are all important places to start to ensure that the transition is unsurprising to existing contributors and users.

When taking over an existing repository, please ensure that the entire Git history of the original repository is kept in the new upstream-maintained repository to make sure the review process isn't stalled during the transition. This is most easily accomplished by forking the new from the existing repository, but can also be accomplished by fetching the commits directly from the original and pushing them into the new repo (ie, git fetch https://github.com/jsmith/example.git master, git rebase FETCH_HEAD, git push -f). On GitHub, an alternative is to move ownership of the git repository. This can be accomplished without giving either group admin access to the other owner's repository:

  • create temporary intermediary organization
  • give old and new owners admin access to intermediary organization
  • old owner transfers repo ownership to intermediary organization
  • new owner transfers repo ownership to its new home
    • recommend that old owner does not fork new repo back into the old organization to ensure that GitHub redirects will just work

Repeatability

Rebuilding the same Dockerfile should result in the same version of the image being packaged, even if the second build happens several versions later, or the build should fail outright, such that an inadvertent rebuild of a Dockerfile tagged as 0.1.0 doesn't end up containing 0.2.3. For example, if using apt to install the main program for the image, be sure to pin it to a specific version (ex: ... apt-get install -y my-package=0.1.0 ...). For dependent packages installed by apt there is not usually a need to pin them to a version.

No official images can be derived from, or depend on, non-official images with the following notable exceptions:

Consistency

All official images should provide a consistent interface. A beginning user should be able to docker run official-image bash (or sh) without needing to learn about --entrypoint. It is also nice for advanced users to take advantage of entrypoint, so that they can docker run official-image --arg1 --arg2 without having to specify the binary to execute.

  1. If the startup process does not need arguments, just use CMD:

    CMD ["irb"]

  2. If there is initialization that needs to be done on start, like creating the initial database, use an ENTRYPOINT along with CMD:

    ENTRYPOINT ["/docker-entrypoint.sh"]
CMD ["postgres"]

    1. Ensure that docker run official-image bash (or sh) works too. The easiest way is to check for the expected command and if it is something else, just exec "$@" (run whatever was passed, properly keeping the arguments escaped).

      #!/bin/sh
set -e

# this if will check if the first argument is a flag
# but only works if all arguments require a hyphenated flag
# -v; -SL; -f arg; etc will work, but not arg1 arg2
if [ "$#" -eq 0 ] || [ "${1#-}" != "$1" ]; then
    set -- mongod "$@"
fi

# check for the expected command
if [ "$1" = 'mongod' ]; then
    # init db stuff....
    # use gosu (or su-exec) to drop to a non-root user
    exec gosu mongod "$@"
fi

# else default to run whatever the user wanted like "bash" or "sh"
exec "$@"

  3. If the image only contains the main executable and its linked libraries (ie no shell) then it is fine to use the executable as the ENTRYPOINT, since that is the only thing that can run:

    ENTRYPOINT ["fully-static-binary"]
CMD ["--help"]

    The most common indicator of whether this is appropriate is that the image Dockerfile starts with scratch (FROM scratch).

Clarity

Try to make the Dockerfile easy to understand/read. It may be tempting, for the sake of brevity, to put complicated initialization details into a standalone script and merely add a RUN command in the Dockerfile. However, this causes the resulting Dockerfile to be overly opaque, and such Dockerfiles are unlikely to pass review. Instead, it is recommended to put all the commands for initialization into the Dockerfile as appropriate RUN or ENV command combinations. To find good examples, look at the current official images.

Some examples at the time of writing:

init

Following the Docker guidelines it is highly recommended that the resulting image be just one concern per container; predominantly this means just one process per container, so there is no need for a full init system. There are two situations where an init-like process would be helpful for the container. The first being signal handling. If the process launched does not handle SIGTERM by exiting, it will not be killed since it is PID 1 in the container (see "NOTE" at the end of the Foreground section in the docker docs). The second situation would be zombie reaping. If the process spawns child processes and does not properly reap them it will lead to a full process table, which can prevent the whole system from spawning any new processes. For both of these concerns we recommend tini. It is incredibly small, has minimal external dependencies, fills each of these roles, and does only the necessary parts of reaping and signal forwarding.

Be sure to use tini in CMD or ENTRYPOINT as appropriate.

It is best to install tini from a distribution-provided package (ex. apt-get install tini). If tini is not available in your distribution or is too old, here is a snippet of a Dockerfile to add in tini:

# Install tini for signal processing and zombie killing
ENV TINI_VERSION v0.18.0
ENV TINI_SIGN_KEY 595E85A6B1B4779EA4DAAEC70B588DFF0527A9B7
RUN set -eux; \
  wget -O /usr/local/bin/tini "https://github.com/krallin/tini/releases/download/${TINI_VERSION}/tini"; \
  wget -O /usr/local/bin/tini.asc "https://github.com/krallin/tini/releases/download/${TINI_VERSION}/tini.asc"; \
  export GNUPGHOME="$(mktemp -d)"; \
  gpg --batch --keyserver ha.pool.sks-keyservers.net --recv-keys "$TINI_SIGN_KEY"; \
  gpg --batch --verify /usr/local/bin/tini.asc /usr/local/bin/tini; \
  command -v gpgconf && gpgconf --kill all || :; \
  rm -r "$GNUPGHOME" /usr/local/bin/tini.asc; \
  chmod +x /usr/local/bin/tini; \
  tini --version

Cacheability

This is one place that experience ends up trumping documentation for the path to enlightenment, but the following tips might help:

  • Avoid COPY/ADD whenever possible, but when necessary, be as specific as possible (ie, COPY one-file.sh /somewhere/ instead of COPY . /somewhere).

    The reason for this is that the cache for COPY instructions considers file mtime changes to be a cache bust, which can make the cache behavior of COPY unpredictable sometimes, especially when .git is part of what needs to be COPYed (for example).

  • Ensure that lines which are less likely to change come before lines that are more likely to change (with the caveat that each line should generate an image that still runs successfully without assumptions of later lines).

    For example, the line that contains the software version number (ENV MYSOFTWARE_VERSION 4.2) should come after a line that sets up the APT repository .list file (RUN echo 'deb http://example.com/mysoftware/debian some-suite main' > /etc/apt/sources.list.d/mysoftware.list).

Security

Image Build

The Dockerfile should be written to help mitigate interception attacks during build. Our requirements focus on three main objectives: verifying the source, verifying author, and verifying the content; these are respectively accomplished by the following: using https where possible; importing PGP keys with the full fingerprint in the Dockerfile to check signatures; embedding checksums directly in the Dockerfile. All three should be used when possible. Just https and embedded checksum can be used when no signature is published. As a last resort, just an embedded checksum is acceptable if the site doesn't have https available and no signature.

The purpose in recommending the use of https for downloading needed artifacts is that it ensures that the download is from a trusted source which also happens to make interception much more difficult.

The purpose in recommending PGP signature verification is to ensure that only an authorized user published the given artifact. When importing PGP keys, please use the high-availability server pool from sks-keyservers (ha.pool.sks-keyservers.net). While there are often transient failures with servers in this pool, the build servers have a proxy that greatly improves reliability (see the FAQ section on keys and verification).

The purpose in recommending checksum verification is to verify that the artifact is as expected. This ensures that when remote content changes, the Dockerfile also will change and provide a natural docker build cache bust. As a bonus, this also prevents accidentally downloading newer-than-expected artifacts on poorly versioned files.

Below are some examples:

  • Preferred: download over https, PGP key full fingerprint import and asc verification, embedded checksum verified.

    ENV PYTHON_DOWNLOAD_SHA512 (sha512-value-here)
RUN set -eux; \
    curl -fL "https://www.python.org/ftp/python/$PYTHON_VERSION/Python-$PYTHON_VERSION.tar.xz" -o python.tar.xz; \
    curl -fL "https://www.python.org/ftp/python/$PYTHON_VERSION/Python-$PYTHON_VERSION.tar.xz.asc" -o python.tar.xz.asc; \
    export GNUPGHOME="$(mktemp -d)"; \
# gpg: key F73C700D: public key "Larry Hastings <larry@hastings.org>" imported
    gpg --batch --keyserver ha.pool.sks-keyservers.net --recv-keys 97FC712E4C024BBEA48A61ED3A5CA953F73C700D; \
    gpg --batch --verify python.tar.xz.asc python.tar.xz; \
    rm -r "$GNUPGHOME" python.tar.xz.asc; \
    echo "$PYTHON_DOWNLOAD_SHA512 *python.tar.xz" | sha512sum --strict --check; \
    # install

  • Alternate: full key fingerprint imported to apt which will check signatures and checksums when packages are downloaded and installed.

    RUN set -eux; \
    key='A4A9406876FCBD3C456770C88C718D3B5072E1F5'; \
    export GNUPGHOME="$(mktemp -d)"; \
    gpg --batch --keyserver ha.pool.sks-keyservers.net --recv-keys "$key"; \
    gpg --batch --armor --export "$key" > /etc/apt/trusted.gpg.d/mysql.gpg.asc; \
    gpgconf --kill all; \
    rm -rf "$GNUPGHOME"; \
    apt-key list > /dev/null

RUN set -eux; \
    echo "deb http://repo.mysql.com/apt/debian/ stretch mysql-${MYSQL_MAJOR}" > /etc/apt/sources.list.d/mysql.list; \
    apt-get update; \
    apt-get install -y mysql-community-client="${MYSQL_VERSION}" mysql-community-server-core="${MYSQL_VERSION}"; \
    rm -rf /var/lib/apt/lists/*; \
    # ...

    (As a side note, rm -rf /var/lib/apt/lists/* is roughly the opposite of apt-get update -- it ensures that the layer doesn't include the extra ~8MB of APT package list data, and enforces appropriate apt-get update usage.)

  • Less Secure Alternate: embed the checksum into the Dockerfile.

    ENV RUBY_DOWNLOAD_SHA256 (sha256-value-here)
RUN set -eux; \
    curl -fL -o ruby.tar.gz "https://cache.ruby-lang.org/pub/ruby/$RUBY_MAJOR/ruby-$RUBY_VERSION.tar.gz"; \
    echo "$RUBY_DOWNLOAD_SHA256 *ruby.tar.gz" | sha256sum -c --strict --check; \
    # install

  • Unacceptable: download the file over http(s) with no verification.

    RUN curl -fL "https://julialang.s3.amazonaws.com/bin/linux/x64/${JULIA_VERSION%[.-]*}/julia-${JULIA_VERSION}-linux-x86_64.tar.gz" | tar ... \
    # install
Runtime Configuration

By default, Docker containers are executed with reduced privileges: whitelisted Linux capabilities, Control Groups, and a default Seccomp profile (1.10+ w/ host support). Software running in a container may require additional privileges in order to function correctly, and there are a number of command line options to customize container execution. See docker run Reference and Seccomp for Docker for reference.

Official Repositories that require additional privileges should specify the minimal set of command line options for the software to function, and may still be rejected if this introduces significant portability or security issues. In general, --privileged is not allowed, but a combination of --cap-add and --device options may be acceptable. Additionally, --volume can be tricky as there are many host filesystem locations that introduce portability/security issues (e.g. X11 socket).

Security Releases

For image updates which constitute a security fix, there are a few things we recommend to help ensure your update is merged, built, and released as quickly as possible:

  1. Contact us a few days in advance to give us a heads up and a timing estimate (so we can schedule time for the incoming update appropriately).
  2. Include [security] in the title of your pull request (for example, [security] Update FooBar to 1.2.5, 1.3.7, 2.0.1).
  3. Keep the pull request free of changes that are unrelated to the security fix -- we'll still be doing review of the update, but it will be expedited so this will help us help you.
  4. Be active and responsive to comments on the pull request after it's opened (as usual, but even more so if the timing of the release is of importance).

Multiple Architectures

Each repo can specify multiple architectures for any and all tags. If no architecture is specified, images are built in Linux on amd64 (aka x86-64). To specify more or different architectures, use the Architectures field (comma-delimited list, whitespace is trimmed). Valid architectures are found in Bashbrew's oci-platform.go file:

  • amd64
  • arm32v6
  • arm32v7
  • arm64v8
  • i386
  • mips64le
  • ppc64le
  • s390x
  • windows-amd64

The Architectures of any given tag must be a strict subset of the Architectures of the tag it is FROM.

We strongly recommend that most images create a single Dockerfile per entry in the library file that can be used for multiple architectures. This means that each supported architecture will have the same FROM line (e.g. FROM debian:jessie). See golang, docker, haproxy, and php for examples of library files using one Dockerfile per entry and see their respective git repos for example Dockerfiles.

For images that are FROM scratch like debian it will be necessary to have a different Dockerfile and build context in order to ADD architecture specific binaries. Since these images use the same Tags, they need to be in the same entry. Use the architecture specific fields for GitRepo, GitFetch, GitCommit, and Directory, which are the architecture concatenated with hyphen (-) and the field (e.g. arm32v7-GitCommit). Any architecture that does not have an architecture-specific field will use the default field (e.g. no arm32v7-Directory means Directory will be used for arm32v7). See the debian or ubuntu files in the library for examples. The following is an example for hello-world:

Maintainers: Tianon Gravi <admwiggin@gmail.com> (@tianon),
             Joseph Ferguson <yosifkit@gmail.com> (@yosifkit)
GitRepo: https://github.com/docker-library/hello-world.git
GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87

Tags: latest
Architectures: amd64, arm32v5, arm32v7, arm64v8, ppc64le, s390x
# all the same commit; easy for us to generate this way since they could be different
amd64-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
amd64-Directory: amd64/hello-world
arm32v5-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
arm32v5-Directory: arm32v5/hello-world
arm32v7-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
arm32v7-Directory: arm32v7/hello-world
arm64v8-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
arm64v8-Directory: arm64v8/hello-world
ppc64le-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
ppc64le-Directory: ppc64le/hello-world
s390x-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
s390x-Directory: s390x/hello-world

Tags: nanoserver
Architectures: windows-amd64
# if there is only one architecture, you can use the unprefixed fields
Directory: amd64/hello-world/nanoserver
# or use the prefixed versions
windows-amd64-GitCommit: 7d0ee592e4ed60e2da9d59331e16ecdcadc1ed87
Constraints: nanoserver

See the instruction format section for more information on the format of the library file.

Commitment

Proposing a new official image should not be undertaken lightly. We expect and require a commitment to maintain your image (including and especially timely updates as appropriate, as noted above).

Library definition files

The library definition files are plain text files found in the library/ directory of the official-images repository. Each library file controls the current "supported" set of image tags that appear on the Docker Hub description. Tags that are removed from a library file do not get removed from the Docker Hub, so that old versions can continue to be available for use, but are not maintained by upstream or the maintainer of the official image. Tags in the library file are only built through an update to that library file or as a result of its base image being updated (ie, an image FROM debian:jessie would be rebuilt when debian:jessie is built). Only what is in the library file will be rebuilt when a base has updates.

Given this policy, it is worth clarifying a few cases: backfilled versions, release candidates, and continuous integration builds. When a new repository is proposed, it is common to include some older unsupported versions in the initial pull request with the agreement to remove them right after acceptance. Don't confuse this with a comprehensive historical archive which is not the intention. Another common case where the term "supported" is stretched a bit is with release candidates. A release candidate is really just a naming convention for what are expected to be shorter-lived releases, so they are totally acceptable and encouraged. Unlike a release candidate, continuous integration builds which have a fully automated release cycle based on code commits or a regular schedule are not appropriate.

It is highly recommended that you browse some of the existing library/ file contents (and history to get a feel for how they change over time) before creating a new one to become familiar with the prevailing conventions and further help streamline the review process (so that we can focus on content instead of esoteric formatting or tag usage/naming).

Filenames

The filename of a definition file will determine the name of the image repository it creates on the Docker Hub. For example, the library/ubuntu file will create tags in the ubuntu repository.

Tags and aliases

The tags of a repository should reflect upstream's versions or variations. For example, Ubuntu 14.04 is also known as Ubuntu Trusty Tahr, but often as simply Ubuntu Trusty (especially in usage), so ubuntu:14.04 (version number) and ubuntu:trusty (version name) are appropriate aliases for the same image contents. In Docker, the latest tag is a special case, but it's a bit of a misnomer; latest really is the "default" tag. When one does docker run xyz, Docker interprets that to mean docker run xyz:latest. Given that background, no other tag ever contains the string latest, since it's not something users are expected or encouraged to actually type out (ie, xyz:latest should really be used as simply xyz). Put another way, having an alias for the "highest 2.2-series release of XYZ" should be xyz:2.2, not xyz:2.2-latest. Similarly, if there is an Alpine variant of xyz:latest, it should be aliased as xyz:alpine, not xyz:alpine-latest or xyz:latest-alpine.

It is strongly encouraged that version number tags be given aliases which make it easy for the user to stay on the "most recent" release of a particular series. For example, given currently supported XYZ Software versions of 2.3.7 and 2.2.4, suggested aliases would be Tags: 2.3.7, 2.3, 2, latest and Tags: 2.2.4, 2.2, respectively. In this example, the user can use xyz:2.2 to easily use the most recent patch release of the 2.2 series, or xyz:2 if less granularity is needed (Python is a good example of where that's most obviously useful -- python:2 and python:3 are very different, and can be thought of as the latest tag for each of the major release tracks of Python).

As described above, latest is really "default", so the image that it is an alias for should reflect which version or variation of the software users should use if they do not know or do not care which version they use. Using Ubuntu as an example, ubuntu:latest points to the most recent LTS release, given that it is what the majority of users should be using if they know they want Ubuntu but do not know or care which version (especially considering it will be the most "stable" and well-supported release at any given time).

Instruction format

The manifest file format is officially based on RFC 2822, and as such should be familiar to folks who are already familiar with the "headers" of many popular internet protocols/formats such as HTTP or email.

The primary additions are inspired by the way Debian commonly uses 2822 -- namely, lines starting with # are ignored and "entries" are separated by a blank line.

The first entry is the "global" metadata for the image. The only required field in the global entry is Maintainers, whose value is comma-separated in the format of Name <email> (@github) or Name (@github). Any field specified in the global entry will be the default for the rest of the entries and can be overridden in an individual entry.

# this is a comment and will be ignored
Maintainers: John Smith <jsmith@example.com> (@example-jsmith),
             Anne Smith <asmith@example.com> (@example-asmith)
GitRepo: https://github.com/example/docker-example.git
GitCommit: deadbeefdeadbeefdeadbeefdeadbeefdeadbeef

# this is also a comment, and will also be ignored

Tags: 1.2.3, 1.2, 1, latest
Directory: 1

Tags: 2.0-rc1, 2.0-rc, 2-rc, rc
GitRepo: https://github.com/example/docker-example-rc.git
GitFetch: refs/heads/2.0-pre-release
GitCommit: beefdeadbeefdeadbeefdeadbeefdeadbeefdead
Directory: 2

Bashbrew will fetch code out of the Git repository (GitRepo) at the commit specified (GitCommit). If the commit referenced is not available by fetching master of the associated GitRepo, it becomes necessary to supply a value for GitFetch in order to tell Bashbrew what ref to fetch in order to get the commit necessary.

The built image will be tagged as <manifest-filename>:<tag> (ie, library/golang with a Tags value of 1.6, 1, latest will create tags of golang:1.6, golang:1, and golang:latest).

Optionally, if Directory is present, Bashbrew will look for the Dockerfile inside the specified subdirectory instead of at the root (and Directory will be used as the "context" for the build instead of the top-level of the repository).

See the multi-arch section for details on how to specify a different GitRepo, GitFetch, GitCommit, or Directory for a specific architecture.

Creating a new repository

  • Create a new file in the library/ folder. Its name will be the name of your repository on the Hub.
  • Add your tag definitions using the appropriate syntax (see above).
  • Create a pull request adding the file from your forked repository to this one. Please be sure to add details as to what your repository does.

Adding a new tag in an existing repository (that you're the maintainer of)

  • Add your tag definition using the instruction format documented above.
  • Create a pull request from your Git repository to this one. Please be sure to add details about what's new, if possible.

Change to a tag in an existing repository (that you're the maintainer of)

  • Update the relevant tag definition using the instruction format documented above.
  • Create a pull request from your Git repository to this one. Please be sure to add details about what's changed, if possible.

Bashbrew

Bashbrew (bashbrew) is a tool for cloning, building, tagging, and pushing the Docker official images. See the Bashbrew README for more information.