engineering thoughts

Applied Docker: Continuous Integration


  • Docker and Jenkins (or any similar tool) are a perfect match for building continuous integration
  • The Dockerfile empowers developers to manage their own OS package requirements
  • Containers as build artifacts significantly simplifies integration scenarios
  • 12-factor app best practices fall out naturally from this strategy for continuous integration


Make no mistake; Docker is the new hotness, but there aren’t many clear explanations of where Docker fits best with modern web stacks today. Frequently I’ll encounter comments along the lines of “Disclaimer: I’ve never tried using Docker, but have you tried it yet?” or “I’m no expert on LXC or docker (never used it, but I’ve read a lot about it) but …” on Hacker News.

Some of my close friends in the industry have shared similar thoughts - in a world where best practices around configuration management, deployment, and virtualization are well-established, why is Docker interesting?

Like any other tool, Docker offers a solution to a specific class of challenges; in particular, Docker exposes a Git-like interface to building and executing lightweight LXC containers. One direct application of Linux containers is powering Heroku-style PaaS runtime environments - it’s a natural extension of the Docker project’s history.

However, most of us aren’t building Heroku knockoffs (nor is it clear that deploying your production application via a git push prod master ad-hoc strategy is a good idea). Since most companies are not Heroku or Google, where else can Docker provide value in cutting edge web architectures? Right now, there aren’t many resources online showing “real-life” applied Docker use cases for the benefit of others.

In this post, I’d like to describe my recent implementation of continuous integration at Standard Treasury as an example of Docker’s strengths. Our stack is a traditional Flask/SQLAlchemy web architecture, with dependencies on Postgres and RabbitMQ - fairly typical of many teams, though the content of this post is applicable regardless of language or framework.


The continuous integration flow I’ve selected works like this:

  • Build a Docker container and mount the repo inside (via ADD in a Dockerfile)
  • Execute tests inside the container, starting & linking external resources as needed
  • Release build artifacts - this can mean one of two paths.
    • For shared libraries, push to private PyPI index, commit a git tag, etc.
    • For web applications, tag the container and push to private Docker repository
  • Clean up temporary containers, remove docker images

My strategy for continuous integration is to isolate test runs by building a new container for each commit. I’ll be using Jenkins in the examples here, but any build tool that can access your git repository and execute docker commands via sudo should work with this approach. Each repository that I’ll be testing will include its own Dockerfile, which will install shared libraries, 3rd-party dependencies, and otherwise perform system setup required for test runs to pass.

For tests that require an external resource (database, redis, cassandra, etc), the build tool will start a separate Docker container and link it to the test execution container. Linking is a recent feature added to Docker in the 0.6.5 release on October 31. Some advantages of launching external services as containers include shared-nothing service isolation and simplicity starting / destroying services.

I’ve opted to use a wrapper script called in each repo to take variables exposed by linked containers and create new ones that the application expects, exec‘ing the remaining arguemnts. The Dockerfile for each application sets the ENTRYPOINT as


The build step is a single shell command:

sudo docker build -t $JOB_NAME/$BUILD_NUMBER .

Each repo contains a top-level Dockerfile that specifies all the steps required to provision a Linux OS from a vanilla image. Since each step is cached, each test run after the initial one should be very fast. Since we have lots of repositories, I’ve opted to create a base python container that each application inherits (via the FROM Dockerfile directive).

Base python Dockerfile

The application-specific Dockerfile is largely uncached, since every RUN step after the ADD will need to execute for each test run.

Application-specific Dockerfile

The ENTRYPOINT script in this example is used to map the Docker-exposed environment variables to the values our application expects. I’ve covered this in detail in my previous blog post.


The test step is also a short shell script:

DB_CONTAINER=$(sudo docker run -d -name $DB_NAME <private repository URL>/database-schema)
sudo docker run -link $DB_NAME:db -t $JOB_NAME/$BUILD_NUMBER nosetests

In this example, the database-schema container represents the latest tagged Postgres schema from a previous test run (against the repository that tracks our schema changes). In this way, tests can layer on top of each other to ensure commits integrate properly with other repositories.


As mentioned earlier, the release step either tags the release via git or tags the container as the latest stable release container, pushing the tag to the remote private docker repository.

sudo docker tag $JOB_NAME/$BUILD_NUMBER <private repository URL>/${JOB_NAME}-master
sudo docker push <private repository URL>/${JOB_NAME}-master >/dev/null


Not much interesting in this step, but you’ll want to docker kill running containers needed only for the test run, docker rm killed containers, and docker rmi the built image.

sudo docker kill ...
sudo docker rm ...
sudo docker rmi ...


As a non-1.0 tool, Docker is still a young project with a lot of potential. Testing infrastructure is a great place to dip your toe into the LXC ecosystem in a low-risk, high-value way. The testing strategy discussed in this post capitalizes on the best of what Docker offers.

The final complete testing script that you can add as a shell script to Jenkins is this:

Full Jenkins test script

Another nice feature of this approach is that it enforces best practices from the 12 factor application manifesto. For your tests to connect to the linked services, environment variables must be used to configure your application. This addresses both config and backing services recommendations. If you chose to run the tested container in production, you’ll be far along the path towards meeting the requirements of dev/prod parity. Even if you opt not to use docker in production, the interface docker enforces will be an asset regardless of deployment technique.


  • Use a wrapper shell script as your Dockerfile ENTRYPOINT to manipulate environment variables exposed via -link‘ed containers.
  • Push service containers into a private Docker repository
  • Tag stable containers once tests pass successfully
  • For integration testing, use the latest stable tagged containers (and make assertions as needed)

Further reading

comments powered by Disqus