Python Base Images

Official Docker Community Images

The Docker “Official Images” for Python 3 based on debian:stretch-slim (55.3 MiB) are around 140 MiB in size. Namely, python:3.6-slim-stretch comes in at 137.9 MiB, while Python 3.7 is a little bigger at 142.7 MiB. The ones based on Alpine have 74.2 MiB and 78.1 MiB. Google’s distroless gcr.io/distroless/python3 beats that with 51 MiB.

The Dockerfiles and related sources for the images can be found at docker-library/python.

Ubuntu Bionic + Python 3

Python 3.6 (Ubuntu default)

This shows how to create a Python 3 image based on Ubuntu Bionic. It is roughly 55 MiB larger than the Alpine-based equivalent, but also comes with runtime essentials like enabled locale, CA certificates, and glibc instead of the often problematic musl libc.

The new ‘minimized’ Ubuntu images are a good base when you want to stay in known waters, and your payload is not just a trivial script or something equally simple. Alpine is a good choice for these small payloads, which also often just have pure-Python dependencies, avoiding most trouble with the different libc.

There is a simple and an optimized version of the Dockerfile, with a few magic incantations added to the latter. To build both versions with clean caches and timings for each build, use this:

docker system prune --all --force
( cd biopy3 \
    && docker pull ubuntu:bionic \
    && time docker build -f Dockerfile.simple -t biopy3-simple . \
    && time docker build -f Dockerfile.optimized -t biopy3 . \
)

Both images are built in about 15 seconds (on an Intel Core i7-6700 with SSD storage). The ‘optimized’ one has a sub-second advantage regarding build time. If you install more packages, the difference should get more pronounced.

The image sizes show a much clearer picture: 129.6 MiB for the ‘optimized’ version, and 176.5 MiB for the ‘simple’ one.

Differences between simple and optimized Dockerfile versions

Differences between simple and optimized Dockerfile versions

Here are the objectives for each of the changes as shown above:

  • -o Acquire::Languages=none speeds up package list downloads by ignoring unneeded translation files.
  • --no-install-recommends limits the installed package set to what you listed explicitly, and hard dependencies of that list – e.g. nodejs will otherwise install a full Python 2.7 for no good reason, instead of just python-minimal. That improves both build times and image size.
  • -o Dpkg::Options::=--force-unsafe-io switches off sync system calls during package expansion, speeding up package installation – since data is saved to a container layer shortly afterwards anyway, this is safe despite the option’s name. ☺
  • apt-get clean && rm -rf "/var/lib/apt/lists"/* removes any cached packages and metadata before the layer is stored. Both are things that we simply do not need in an immutable container. While the APT clean is already baked into newer Debian base images, having an explicit cleanup call doesn’t hurt either.

And the env LANG=C before the apt-get commands suppresses locale initialization warnings since locales are not generated yet.

Python 3.7 (Deadsnakes PPA)

In biopy3/Dockerfile.deadsnakes the newest Python version is added, as available from the Deadsnakes PPA.

Due to packaging mechanics, this gets installed in addition to Ubuntu’s default Python 3.6 – the resulting image size is 168.9 MiB (i.e. ~40 MiB more). That means this is not a sensible option compared to images like python:3.7-slim-stretch. Also, timely security updates are not guaranteed for the PPA release channel.

Note that for the pip installation via get-pip.py, the command set -o pipefail is used to ensure the build fails if wget fails. That in turn requires using the SHELL instruction to switch the default shell from dash, which does not implement set -o shell options, to bash which supports that option.

SHELL ["/bin/bash", "-c"]
…
RUN …
    && set -o pipefail \
    && wget -qO- https://bootstrap.pypa.io/get-pip.py | python3.7 \
    …

Python 3.7 (pyenv)

To build images for Python 3.7 and up, compiled and installed using pyenv, run these commands:

command cd $(git rev-parse --show-toplevel)/biopy3
pyversion=$(grep ARG.pyversion= Dockerfile.pyenv | cut -f2 -d= | cut -f1-2 -d.)
pyenv_version=$(grep ARG.pyenv_version= Dockerfile.pyenv | cut -f2 -d=)
wget https://github.com/pyenv/pyenv/archive/v${pyenv_version}.tar.gz
./make-pyenv-tk.sh >/dev/null

declare -A base_images=( [biopyenv3]="ubuntu:bionic" [debpyenv3]="debian:stretch-slim" )
for tag in "${!base_images[@]}"; do
    docker build --build-argument distro="${base_images[$tag]}" \
                 -t ${tag} -t ${tag}:$pyversion -f Dockerfile.pyenv .
    docker build --build-argument distro="${base_images[$tag]}" \
                 -t ${tag}-tk -t ${tag}-tk:$pyversion -f Dockerfile.pyenv-tk .
done

See the comments at the start of Dockerfile.pyenv for some more details and resulting image sizes.

TODO

Look at other options like Conda or PyRun.