Multi-Stage Dockerfiles and Python Virtualenvs

Using Docker’s multi-stage build feature and Python’s virtualenv tool, we can make smaller and more secure docker images for production.

The Problem

We want to have the smallest production Docker images for a number of reasons. Having files unnecessary to the running app makes the image slower to move to and from the Docker Hub, as well as potentially introducing security vulnerabilities. So we’d really prefer not to have any of the build tools or libraries required to build the app in the final image if we can help it.

To accomplish this we used to have several docker files and we’d build them one at a time. We’d then either copy the built files out of them, or mount a volume and run the process that produced the files we needed. This was cumbersome however, and could potentially conflict with other builds on the same CI box if we weren’t careful. This technique, dubbed the “Builder Pattern”, did work well for a time, but when we saw the ability to get the same benefit in a single build step we jumped at the chance.

Multi-Stage Dockerfile

A multi-stage Dockerfile is one that is basically more than one Dockerfile glued together; i.e. it has more than one FROM line. The resulting image is defined in the last section in the file. So, for example, you could base the first section of the file on a NodeJS base image, and in there have it build and process your static assets (CSS, JS, images). Then in a second section based on a Python image you build your app and copy the assets from the first stage. Here’s a simple example:

# assets builder and dev server
FROM node:8 AS assets

# Add node tools to path
ENV PATH=/app/node_modules/.bin:$PATH

# copy dependency definitions
COPY package.json yarn.lock ./

# install dependencies
RUN yarn install --pure-lockfile
RUN yarn global add gulp-cli@2.0.1

# copy supporting files and media
COPY .eslintrc.js .stylelintrc gulpfile.js ./
COPY ./media ./media

# build production files
RUN gulp build --production

The above results in processed files in an static_final directory in the image. Another section can get those files with a directive like COPY --from=assets /app/static_final /app/static_final (see below).

Python Virtualenv

Normally in a Docker image a virtualenv is superfluous. You can just install your Python dependencies globally in the image and it won’t interfere with a thing (Yay Docker!). But the multi-stage dockerfile (or just using multiple dockerfiles before this feature) introduces a reason to use them again. We can use a larger image with a lot of build tools and dependencies to compile and install all of our packages in a virtualenv, and then (as long as you keep it at the same path and on the same base distro) copy it to a much leaner and more secure production image. We’d used this builder technique for static assets for a while as mentioned above, but using it for the Python environment I thought was a stroke of genius.

This is a technique I learned from my friend and colleague Giorgos in his work on SUMO at Mozilla. He is awesome and you should check out his stuff.

The basic pattern is as follows:

# Python dependencies builder
# Full official Debian-based Python image
FROM python:3-stretch AS builder

# Always set a working directory
# Sets utf-8 encoding for Python et al
# Turns off writing .pyc files; superfluous on an ephemeral container.
# Seems to speed things up

# Ensures that the python and pip executables used
# in the image will be those from our virtualenv.
ENV PATH="/venv/bin:$PATH"

# Install OS package dependencies.
# Do all of this in one RUN to limit final image size.
RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        gettext build-essential mariadb-client libmariadbclient-dev \
        libxml2-dev libxslt1-dev libxslt1.1 && \
    rm -rf /var/lib/apt/lists/*

# Setup the virtualenv
RUN python -m venv /venv
# or "virtualenv /venv" for Python 2

# Install Python deps
COPY requirements.txt ./
RUN pip install --no-cache-dir -r requirements.txt

And then the production app image:

# django app container
# Smaller official Debian-based Python image
FROM python:3-slim-stretch AS app

# Extra python env
ENV PATH="/venv/bin:$PATH"

CMD ["./bin/"]

# do all of this in one RUN to limit final image size
RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        gettext libxslt1.1 && \
    rm -rf /var/lib/apt/lists/*

# copy in Python environment
COPY --from=builder /venv /venv

# copy in static assets
COPY --from=assets /app/static_final /app/static_final

# copy in the rest of the app
COPY ./ ./

You could just copy the entire Python environment from one to the other, but that is a lot more to copy (includes the stdlib) and in my opinion more cumbersome. A virtualenv is an already convenient and portable Python environment.

The final Dockerfile is just the three above sections concatenated together. The final image has no extra build tools, language runtimes (Node in this case), or libraries required only during build. In the end the container running your service is much smaller and more secure.

See the official Python docker image repository for descriptions of the image types if you’re curious. I highly recommend the official Python images. They are well maintained and always have the latest Python versions.

Wrapping Up

I really like this technique. You get the benefit of a single Dockerfile with the full build encapsulated into a single file, while also producing a much smaller and more secure image. I’ve thrown together a quick demo app repo so that you can see it all really work and fit together, and because I wanted an excuse to play with Responder :)

I also included in the demo, simple as it is, an example Dockerfile of the same commands but just using one build stage. The results of building these are:

$ docker images | grep multi-stage
multi-stage-docker-venv-demo_app         latest  a6106f3a4c59  195MB
multi-stage-docker-venv-demo_app-pipenv  latest  e945e860ec24  216MB
multi-stage-docker-venv-demo_app-full    latest  e945e860ec24  979MB

I hope you found this helpful. Thanks for reading!

Improve this page