tree: a0af83da18553625fbb59ca0b9511211fc88a9dc [path history] [tgz]
  1. resources/
  2. __init__.py
  3. __main__.py
  4. builder.py
  5. cipd.py
  6. dockcross.py
  7. dockerbuild.py
  8. markdown.py
  9. platform.py
  10. README.md
  11. README.wheels.md
  12. runtime.py
  13. source.py
  14. standalone.vpython
  15. types.py
  16. util.py
  17. wheel.py
  18. wheel_cryptography.py
  19. wheel_infra.py
  20. wheel_mysql.py
  21. wheel_opencv.py
  22. wheel_wheel.py
  23. wheels.md
infra/tools/dockerbuild/README.md

DockerBuild

infra.tools.dockerbuild is a management and execution environment for Chrome Infrastructure's Docker toolchain build environment. Fundamentally, it handles:

  • Generation and internment of Chrome Infrastructure Docker build images.
  • Specific build instructions for Python wheels.
  • Entry point for build operations using Docker toolchains.

DockerBuild is complex. It justifies this complexity by solving some annoying problems:

  • Cross-compiling is really useful. It is difficult to have continuous builders for each supported platform, especially the more esoteric ones. Adding support for a new platform is decoupled from actually having that platform running.
  • Cross-compiling Python wheels is nuanced and full of errors. Python build scripts, in general, do not really accommodate cross-compiling without hacks. These hacks, and the overall logic, should be enshrined somewhere.
  • Building the Python cryptography wheel for Linux really sucks. This logic and all of its prerequisites should be enshrined in a script somewhere.
  • In general, it is really good to zip up build toolchains and environments and use them hermetically instead of relying on the configuration of a builder somewhere.

DockerBuild Images

DockerBuild images are based on dockcross project images. Each dockcross image bundles a platform‘s toolchain with an entry point script which can be used to execute commands within that toolchain. The script, when run, mounts a directory to /work within the Docker image and invokes the provided command on that directory. This allows a user to almost-seamlessly use the dockcross image’s toolchain to operate on local files. This, in turn makes it really easy to cross-compile.

DockerBuild extends dockcross images to add:

  • A host and cross-compiled Python environment.
  • A host Perl installation.
  • Several packages in support of generation of complex Python wheels.

See dockcross.py for more details.

DockerBuild images, like dockcross images, can be run using a generated entry point script. This can be created by running the image:

docker run <image-name> > ./entrypoint.sh
chmod +x ./entrypoint.sh
./entrypoint.sh

Alternatively, the script can be run through dockerbuild using the run subcommand.

Note on Windows and Mac

It is important to note that Docker image environments only exist for Linux systems at the moment. Chrome Infrastructure's build strategy for Windows and Mac is to actually perform the builds on real Windows and Mac systems. Support for Windows and Mac Python wheels is built into DockerBuild, but serves mostly as a wrapper around using pip to download pre-generated packages and upload them to CIPD.

Windows and Mac support are gated on the ability to have a fully-functioning Windows and Mac build toolchain running in a Linux (i.e., Docker) environment. This may be possible with a combination of wine and/or a Mac cross-compiler, but this is not currently explored or implemented.

Lifecycle

Walking backwards through the lifecycle of this script:

  • Building Python wheels requires
  • Python wheel source files, which requires
  • Downloading and storing Python sources in CIPD.
  • It also requires a DockerBuild image, which requires
  • A dockcross base image and a set of sources.
  • The sources need to be interned in CIPD for consistency.
  • The dockcross base image needs to be built independently.

Minimal operation will use cached pre-built DockerBuild images and cached CIPD sources, allowing a user to avoid building everything from scratch.

However, the logic to build everything from scratch is included in DockerBuild for updates and reproducibility.

Subcommand: sources

The various sources used to build Python wheels and DockerBuild images must be downloaded from the Internet and interned in CIPD packages. This ensures that any given image will be reproducible and resilient to external packaging and infrastructure changes.

The sources subcommand scans through all sources used by DockerBuild, acquires them, and creates CIPD packages for them. Run sources with the --help flag for more information.

If new sources are added, or sources without CIPD packages are identified, a warning message will be printed at the end of DockerBuild operation encouraging the user to upload source CIPD packages for future reproducible builds.

Subcommand: docker-mirror

The dockcross project‘s images are used as base images for DockerBuild’s specialized images. Rather than rely on upstream Docker hosts, DockerBuild offers the docker-mirror subcommand to load upstream dockcross images and store them in our Google Cloud Docker image registry.

docker-mirror can be run to synchronize and/or update the local images. All DockerBuild images are generated from the local images, not the upstream images, so an explicit synchronization step is required to update their dockcross bases.

Subcommand: docker-generate

DockerBuild's specialized Docker images can be built locally using the docker-generate subcommand. This will pull the base dockcross images and required sources and construct local DockerBuild images.

Optionally, the local images can be pushed to the Google Cloud Docker image registry using the --upload flag. This ensures that other users of the script can retrieve and use the images without requiring them to build them locally.

Subcommand: wheel-build

DockerBuild has specific support for building Python wheels and uploading them to CIPD. This synergizes with vpython's CIPD wheel packaging expectations.

Wheels can be built and uploaded by using the wheel-build subcommand, and uploaded using the --upload flag.

Support for additional Python wheels and/or wheel versions can be added by editing wheel.py and adding entries for the new wheels.

Adding support for universal wheels is easy. Adding support for most platform-specific binary wheels can range from easy to hacky-difficult depending on the wheel's binary requirements and suitability for cross-compiling. This has to be evaluated on a case-by-base basis, sadly.

Subcommand: run

DockerBuild offers an entry point into a DockerBuild image's toolchain environment through its run subcommand. run will mount the specified directory inside of the DockerBuild environment and execute the specified command against that directory.

run can be used to cross-compile software (e.g., git) for other Infra platforms.

Problems with Known Fixes

Got permission denied while trying to connect to the Docker daemon socket...

Fix: Add your local user to the docker POSIX group:

sudo usermod -a -G docker $USER

You need to restart your shell for this to take effect.

Got permission denied while trying to upload CIPD package...

Upload rights are restricted to troopers and members of vpython-wheels-uploaders

First, make sure you are logged into cipd with your @google.com account.

cipd auth-info
cipd auth-login

Second, file a bug with the Infra>Platform>Admin component to add you to this group.

Got permission denied while downloading docker image

When running commands that require docker images, such as the docker-generate subcommand, you need to be authenticated. It is recommended to use gcloud to handle this; try running gcloud auth configure-docker. Authenticating with a google.com account using gcloud auth login may be necessary.

Error response from daemon: squash is only supported with experimental mode

Fix: Run docker in experimental mode (yes, this is a hack):

vim /etc/default/docker
# Edit file such that DOCKER_OPTS includes "--experimental=true"
sudo service docker restart

Examples

Upload all sources to CIPD

To upload all sources to CIPD, run:

./run.py infra.tools.dockerbuild --upload-sources sources

Update dockcross base image mirrors

To update the Google Cloud repository's mirrored dockcross images, run:

./run.py infra.tools.dockerbuild docker-mirror --upload

Build and upload DockerBuild images

To build new DockerBuild images and upload them to the Google Cloud repository, run:

./run.py infra.tools.dockerbuild docker-generate --upload

Regenerate all configured wheels

To ensure that all configured Python wheels are uploaded to CIPD for all known platforms, run the following command:

./run.py infra.tools.dockerbuild --upload-sources wheel-build --upload
  • --upload-sources instructs DockerBuild to upload CIPD packages for any new wheel sources that are not already locally mirrored.
  • --upload instructs the wheel-build subcommand to upload any wheels that are not currently present in CIPD.

This can be run after adding a new wheel configuration, or after adding a new platform to support.

Cross-compile Python

To cross-compile Python for the “linux-armv6” platform, download the Python source, then configure it for the cross-compile environment and build.

wget https://github.com/python/cpython/archive/v2.7.13.tar.gz
cd cpython-2.7.13
./run.py infra.tools.dockerbuild run --platform linux-armv6 run -- \
  sh -c './configure --prefix=/work/PREFIX --host=$CROSS_TRIPLE --build=$(gcc -dumpmachine)'
./run.py infra.tools.dockerbuild run --platform linux-armv6 run -- make install
  • Note that we use sh to ensure that the environment variables are evaluated within the container.
  • Note that the install prefix is /work/PREFIX. This will install into the working directory, but all paths will have /work/PREFIX hard-coded as their prefix. More specific ./configure options can be used to ensure that the configured environment matches the target system environment.