Conda Docker

Conda Docker is an extension to the docker concept by having declarative environments that are associated with docker images. In addition this tool does not require docker to build images. Thus allowing for interesting caching behavior and tricks that docker would not normally allow.

Features:

• docker is not needed to build images
• declarative environments that map 1:1 to docker images
• significantly faster build times since builds can take advantage of package cache
• interesting opportunities for layering (e.g. mkl gets separate layer)
• no dependencies allowing for library packaged as zipapp

Quickstart

Build conda docker image from command line:

conda docker build -b frolvlad/alpine-glibc:latest \
                   -i example-image:123456 \
                   -o demo.tar \
                   numpy numba flask

Examples using Library

Downloading docker images without docker!

from conda_docker.registry.client import pull_image

image = pull_image('frolvlad/alpine-glibc', 'latest')

Modify docker image from filesystem

from conda_docker.docker.base import Image
from conda_docker.registry.client import pull_image

image = pull_image('continuumio/miniconda3', 'latest')
image.remove_layer()
image.name = 'this-is-a-test'
image.add_layer_path('./')
image.add_layer_contents({
    'this/is/a/test1': b'this is test 1',
    'this/is/a/test2': b'this is test 2'
})
image.layers[0].config['Env'].append('FOO=BAR')
image.write_file('example-filter.tar')

Build conda docker image from library

from conda_docker.conda import build_docker_environment

build_docker_environment(
    base_image='frolvlad/alpine-glibc:latest',
    output_image='example-image:123456',
    packages=[
        'numpy',
        'numba',
        'flask',
    ],
    output_filename='demo.tar')

How does this work?

Turns out that docker images are just a tar collection of files. There are several versions of the spec. For v1.0 the specification is defined here. Instead of writing down the spec lets look into a single docker image.

docker pull ubuntu:latest
docker save ubuntu:latest -o /tmp/ubuntu.tar

List the directory structure of the docker image. Notice how it is a collection of layer.tar which is a tar archive of filesystems. And several json files. VERSION file is always 1.0 currently.

tar -tvf /tmp/ubuntu.tar

Dockerhub happens to export docker images in a v1 - v1.2 compatible format. Lets only look at the files important for v1. Repositories tells the layer to use as the layer head of the current name/tag.

tar -xf /tmp/ubuntu.tar $filename
cat $filename | python -m json.tool

For each layer there are three files: VERSION, layer.tar, and json.

tar -xf /tmp/ubuntu.tar $filename
cat $filename

tar -xf /tmp/ubuntu.tar $filename
cat $filename | python -m json.tool

Looking at layer metadata.

{
    "id": "93935bf1450219e4351893e546b97b4584083b01d19daeba56cab906fc75fc1c",
    "created": "1969-12-31T19:00:00-05:00",
    "container_config": {
        "Hostname": "",
        "Domainname": "",
        "User": "",
        "AttachStdin": false,
        "AttachStdout": false,
        "AttachStderr": false,
        "Tty": false,
        "OpenStdin": false,
        "StdinOnce": false,
        "Env": null,
        "Cmd": null,
        "Image": "",
        "Volumes": null,
        "WorkingDir": "",
        "Entrypoint": null,
        "OnBuild": null,
        "Labels": null
    },
    "os": "linux"
}

Looking at the layer filesystem.

tar -xf /tmp/ubuntu.tar $filename
tar -tvf $filename | head

References

• Docker Registry API Specification
• Docker Image Specification
  • Summary
  • Registry V2 Specification