Container environments

The Beam SDK runtime environment is isolated from other runtime systems because the SDK runtime environment is containerized with Docker. This means that any execution engine can run the Beam SDK.

This page describes how to customize, build, and push Beam SDK container images.

Before you begin, install Docker on your workstation.

Customizing container images

You can add extra dependencies to container images so that you don’t have to supply the dependencies to execution engines.

To customize a container image, either:

It’s often easier to write a new Dockerfile. However, by modifying the original Dockerfile, you can customize anything (including the base OS).

Writing new Dockerfiles on top of the original

  1. Pull a prebuilt SDK container image for your target language and version. The following example pulls the latest Python SDK:
    docker pull apachebeam/python3.7_sdk
  2. Write a new Dockerfile that designates the original as its parent.
  3. Build a child image.

Modifying the original Dockerfile

  1. Clone the beam repository:
    git clone
  2. Customize the Dockerfile. If you’re adding dependencies from PyPI, use base_image_requirements.txt instead.
  3. Reimage the container.

Testing customized images

To test a customized image locally, run a pipeline with PortableRunner and set the --environment_config flag to the image path:

python -m apache_beam.examples.wordcount \
--input=/path/to/inputfile \
--output /path/to/write/counts \
--runner=PortableRunner \
--job_endpoint=embed \
# Start a Spark job server on localhost:8099
./gradlew :runners:spark:job-server:runShadow

# Run a pipeline on the Spark job server
python -m apache_beam.examples.wordcount \
--input=/path/to/inputfile \
--output=path/to/write/counts \
--runner=PortableRunner \
--job_endpoint=localhost:8099 \

To test a customized image on the Google Cloud Dataflow runner, use DataflowRunner with the beam_fn_api experiment and set worker_harness_container_image to the custom container:

python -m apache_beam.examples.wordcount \ 
--input=path/to/inputfile \
--output=/path/to/write/counts \
--runner=DataflowRunner \
--project={gcp_project_id} \
--temp_location={gcs_location} \ \
--experiment=beam_fn_api \
--sdk_location=[…]/beam/sdks/python/container/py{version}/build/target/apache-beam.tar.gz \

# The sdk_location option accepts four Python version variables: 2, 35, 36, and 37

Building container images

To build Beam SDK container images:

  1. Navigate to the local copy of your customized container image.
  2. Run Gradle with the docker target. If you’re building a child image, set the optional --file flag to the new Dockerfile. If you’re building an image from an original Dockerfile, ignore the --file flag and use a default repository:
# The default repository of each SDK
./gradlew [--file=path/to/new/Dockerfile] :sdks:java:container:docker
./gradlew [--file=path/to/new/Dockerfile] :sdks:go:container:docker
./gradlew [--file=path/to/new/Dockerfile] :sdks:python:container:py2:docker
./gradlew [--file=path/to/new/Dockerfile] :sdks:python:container:py35:docker
./gradlew [--file=path/to/new/Dockerfile] :sdks:python:container:py36:docker
./gradlew [--file=path/to/new/Dockerfile] :sdks:python:container:py37:docker

# Shortcut for building all four Python SDKs
./gradlew [--file=path/to/new/Dockerfile] :sdks:python:container buildAll

To examine the containers that you built, run docker images from anywhere in the command line. If you successfully built all of the container images, the command prints a table like the following:

REPOSITORY                          TAG                 IMAGE ID            CREATED           SIZE
apachebeam/java_sdk                 latest              16ca619d489e        2 weeks ago        550MB
apachebeam/python2.7_sdk            latest              b6fb40539c29        2 weeks ago       1.78GB
apachebeam/python3.5_sdk            latest              bae309000d09        2 weeks ago       1.85GB
apachebeam/python3.6_sdk            latest              42faad307d1a        2 weeks ago       1.86GB
apachebeam/python3.7_sdk            latest              18267df54139        2 weeks ago       1.86GB
apachebeam/go_sdk                   latest              30cf602e9763        2 weeks ago        124MB

Overriding default Docker targets

The default tag is latest and the default repositories are in the Docker Hub apachebeam namespace. The docker command-line tool implicitly pushes container images to this location.

To tag a local image, set the docker-tag option when building the container. The following command tags a Python SDK image with a date.

./gradlew :sdks:python:container:py2:docker -Pdocker-tag=2019-10-04

To change the repository, set the docker-repository-root option to a new location. The following command sets the docker-repository-root to a Bintray repository named apache.

./gradlew :sdks:python:container:py2:docker -Pdocker-repository-root=$

Pushing container images

After building a container image, you can store it in a remote Docker repository.

The following steps push a Python SDK image to the docker-root-repository value.

  1. Sign in to your Docker registry:
    docker login
  2. Navigate to the local copy of your container image and upload it to the remote repository:
    docker push apachebeam/python2.7_sdk

To download the image again, run docker pull:

docker pull apachebeam/python2.7_sdk

Note: After pushing a container image, the remote image ID and digest match the local image ID and digest.