Beam code change guide

Last Updated: Apr 18, 2024

This guide is for Beam users and developers who want to change or test Beam code. Specifically, this guide provides information about:

Testing code changes locally
Building Beam artifacts with modified Beam code and using the modified code for pipelines

The guide contains the following sections:

Repository structure: A description of the Apache Beam GitHub repository, including steps for setting up your Gradle project and for verifying the configuration of your develoment environment.
Java Guide: Guidance for setting up a Java environment, running and writing integration tests, and running a pipeline with modified Beam code.
Python Guide: Guidance for configuring your console for Python development, running unit and integration tests, and running a pipeline with modified Beam code.

Repository structure

The Apache Beam GitHub repository (Beam repo) is, for the most part, a “mono repo”. It contains everything in the Beam project, including the SDK, test infrastructure, dashboards, the Beam website, and the Beam Playground.

Code paths

The following example code paths in the Beam repo are relevant for SDK development.

Java

Java code paths are mainly found in two directories: sdks/java and runners. The following list provides notes about the contents of these directories and some of the subdirectories.

sdks/java - Java SDK
- sdks/java/core - Java core
- sdks/java/harness - SDK harness (entrypoint of SDK container)
runners - Java runner supports, including the following items:
- runners/direct-java - Java direct runner
- runners/flink-java - Java Flink runner
- runners/google-cloud-dataflow-java - Dataflow runner (job submission, translation, and so on)
  - runners/google-cloud-dataflow-java/worker - Worker for Dataflow jobs that don't use Runner v2

Non-Java SDKs

For SDKs in other languages, the sdks/LANG directory contains the relevant files. The following list provides notes about the contents of some of the subdirectories.

sdks/python - Setup file and scripts to trigger test-suites
- sdks/python/apache_beam - The Beam package
  - sdks/python/apache_beam/runners/worker - SDK worker harness entrypoint and state sampler
  - sdks/python/apache_beam/io - I/O connectors
  - sdks/python/apache_beam/transforms - Most core components
  - sdks/python/apache_beam/ml - Beam ML code
  - sdks/python/apache_beam/runners - Runner implementations and wrappers
  - ...
sdks/go - Go SDK
.github/workflow - GitHub action workflows, such as the tests that run during a pull request. Most workflows run a single Gradle command. To learn which command to run locally during development, during tests, check which command is running.

Gradle quick start

The Beam repo is a single Gradle project that contains all components, including Java, Python, Go, and the website. Before you begin development, familiarize yourself with the Gradle project structure by reviewing Structuring Projects with Gradle in the Gradle documentation.

Gradle key concepts

Grade uses the following key concepts:

project: a folder that contains the build.gradle file
task: an action defined in the build.gradle file
plugin: predefined tasks and hierarchies; runs in the project's build.gradle file

Common tasks for a Java project or subproject include:

compileJava - compiles the Java source files
compileTestJava - compiles the Java test source files
test - runs unit tests
integrationTest - runs integration tests

To run a Gradle task, use the command ./gradlew -p <PROJECT_PATH> <TASK> or the command ./gradlew :<PROJECT>:<PATH>:<TASK>. For example:

./gradlew -p sdks/java/core compileJava

./gradlew :sdks:java:harness:test

Beam-specific Gradle project configuration

For Apache Beam, one plugin manages everything: buildSrc/src/main/groovy/org/apache/beam/gradle/BeamModulePlugin. The BeamModulePlugin is used for the following tasks:

Manage Java dependencies
Configure projects such as Java, Python, Go, Proto, Docker, Grpc, and Avro
- For Java, use applyJavaNature; for Python, use applyPythonNature
- Define common custom tasks for each type of project
  - test: run Java unit tests
  - spotlessApply: format Java code

In every Java project or subproject, the build.gradle file starts with the following code:


apply plugin: 'org.apache.beam.module'

applyJavaNature( ... )

Environment setup

To set up a local development environment, first review the Contribution guide. If you plan to use Dataflow, you need to set up gcloud credentials. To set up gcloud credentials, see Create a Dataflow pipeline using Java in the Google Cloud documentation.

Depending on the languages involved, your PATH file needs to have the following elements configured.

A Java environment that uses a supported Java version, preferably Java 8.
- This environment is needed for all development, because Beam is a Gradle project that uses JVM.
- Recommended: To manage Java versions, use sdkman.
A Python environment that uses any supported Python version.
- This environment is needed for Python SDK development.
- Recommended: To manage Python versions, use pyenv and a virtual environment.
A Go environment that uses latest Go version.
- This environment is needed for Go SDK development.
- This environment is also needed for SDK container changes for all SDKs, because the container entrypoint scripts are written in Go.
A Docker environment. This environment is needed for the following tasks:
- SDK container changes.
- Some cross-language functionality (if you run an SDK container image; not required in Beam 2.53.0 and later verions).
- Portable runners, such as using job server.

The following list provides examples of when you need specific environemnts.

When you test the code change in sdks/java/io/google-cloud-platform, you need a Java environment.
When you test the code change in sdks/java/harness, you need a Java environment, a Go environment, and Docker environment. You need the Docker environment to compile and build the Java SDK harness container image.
When you test the code change in sdks/python/apache_beam, you need a Python environment.

Java development guide

This section provides guidance for setting up your environment to modify or test Java code.

IDE (IntelliJ) setup

To set up IntelliJ, follow these steps. The IDE isn't required for changing the code and testing. You can run tests can by using a Gradle command line, as described in the Console setup section.

From IntelliJ, open /beam (Important: Open the repository root directory, not sdks/java).
Wait for indexing. Indexing might take a few minutes.

Because Gradle is a self-contained build tool, if the prerequisites are met, the environment setup is complete.

To verify whether the load is successful, follow these steps:

Find the file examples/java/build.gradle.
Next to the wordCount task, a Run button is present. Click Run. The wordCount example compiles and runs.

Console setup

To run tests by using the Gradle command line (shell), in the command-line environment, run the following command. This command compiles the Apache Beam SDK, the WordCount pipeline, and a Hello-world program for data processing. It then runs the pipeline on the Direct Runner.

$ cd beam
$ ./gradlew :examples:java:wordCount

When the command completes successfully, the following text appears in the Gradle build log:

...
BUILD SUCCESSFUL in 2m 32s
96 actionable tasks: 9 executed, 87 up-to-date
3:41:06 PM: Execution finished 'wordCount'.

In addition, the following text appears in the output file:


$ head /tmp/output.txt*
==> /tmp/output.txt-00000-of-00003 <==
should: 38
bites: 1
depraved: 1
gauntlet: 1
battle: 6
sith: 2
cools: 1
natures: 1
hedge: 1
words: 9

==> /tmp/output.txt-00001-of-00003 <==
elements: 1
Advise: 2
fearful: 2
towards: 4
ready: 8
pared: 1
left: 8
safe: 4
canst: 7
warrant: 2

==> /tmp/output.txt-00002-of-00003 <==
chanced: 1
...

Run a unit test

This section explains how to run unit tests locally after you make a code change in the Java SDK, for example, in sdks/java/io/jdbc.

Tests are stored in the src/test/java folder of each project. Unit tests have the filename .../**Test.java. Integration tests have the filename .../**IT.java.

To run all unit tests under a project, use the following command:
```
./gradlew :sdks:java:harness:test
```
Find the JUnit report in an HTML file in the file path <invoked_project>/build/reports/tests/test/index.html.

To run a specific test, use the following commands:

./gradlew :sdks:java:harness:test --tests org.apache.beam.fn.harness.CachesTest
./gradlew :sdks:java:harness:test --tests *CachesTest
./gradlew :sdks:java:harness:test --tests *CachesTest.testClearableCache

To run tests using IntelliJ, click the ticks to run either a whole test class or a specific test. To debug the test, set breakpoints.
These steps don‘t apply to sdks:java:core tests. To invoke those unit tests, use the command :runners:direct-java:needsRunnerTest. Java core doesn’t depend on a runner. Therefore, unit tests that run a pipeline require the Direct Runner.

To run integration tests, use the Direct Runner.

Run integration tests

Integration tests have the filename .../**IT.java. They use TestPipeline. Set options by using TestPipelineOptions.

Integration tests differ from standard pipelines in the following ways:

By default, they block on run (on TestDataflowRunner).
They have a default timeout of 15 minutes.
The pipeline options are set in the system property beamTestPipelineOptions.

To configure the test, you need to set the property -DbeamTestPipelineOptions=[...]. This property sets the runner that the test uses.

The following example demonstrates how to run an integration test by using the command line. This example includes the options required to run the pipeline on the Dataflow runner.

-DbeamTestPipelineOptions='["--runner=TestDataflowRunner","--project=mygcpproject","--region=us-central1","--stagingLocation=gs://mygcsbucket/path"]'

Write integration tests

To set up a TestPipeline object in an integration test, use the following code:

@Rule public TestPipeline pipelineWrite = TestPipeline.create();

@Test
public void testSomething() {
  pipeline.apply(...);

  pipeline.run().waitUntilFinish();
}

The task that runs the test needs to specify the runner. The following examples demonstrate how to specify the runner:

To run a Google Cloud I/O integration test on the Direct Runner, use the command :sdks:java:io:google-cloud-platform:integrationTest.
To run integration tests on the standard Dataflow runner, use the command :runners:google-cloud-dataflow-java:googleCloudPlatformLegacyWorkerIntegrationTest.
To run integration test on Dataflow runner v2, use the command :runners:google-cloud-dataflow-java:googleCloudPlatformRunnerV2IntegrationTest.

To see how to run your workflow locally, refer to the Gradle command that the GitHub Action workflow runs.

The following commands demonstrate an example invocation:

./gradlew :runners:google-cloud-dataflow-java:examplesJavaRunnerV2IntegrationTest \
-PdisableSpotlessCheck=true -PdisableCheckStyle=true -PskipCheckerFramework \
-PgcpProject=<your_gcp_project> -PgcpRegion=us-central1 \
-PgcsTempRoot=gs://<your_gcs_bucket>/tmp

Run your pipeline with modified beam code

To apply code changes to your pipeline, we recommend that you start with a separate branch.

If you're making a pull request or want to test a change with the dev branch, start from Beam HEAD (master).
If you're making a patch on released Beam (2.xx.0), start from a tag, such as v2.55.0. Then, in the Beam repo, use the following command to compile the project that includes the code change. This example modifies sdks/java/io/kafka.
```
./gradlew -Ppublishing -p sdks/java/io/kafka publishToMavenLocal
```
By default, this command publishes the artifact with modified code to the Maven Local repository (~/.m2/repository). The change is picked up when the user pipeline runs.

If your code change is made in a development branch, such as on Beam master or a PR, the artifact is produced under version 2.xx.0-SNAPSHOT instead of on a release tag. To pick up this dependency, you need to make additional configurations in your pipeline project. The following examples provide guidance for making configurations in Maven and Gradle.

Follow these steps for Maven projects.

Recommended: Use the WordCount maven-archetype as a template to set up your project (https://beam.apache.org/get-started/quickstart-java/).

To add a snapshot repository, include the following elements:

<repository>
  <id>Maven-Snapshot</id>
  <name>maven snapshot repository</name>
  <url>https://repository.apache.org/content/groups/snapshots/</url>
</repository>

In the pom.xml file, modify the value of beam.version:

<properties>
<beam.version>2.XX.0-SNAPSHOT</beam.version>

Follow these steps for Gradle projects.

In the build.gradle file, add the following code:

repositories {
maven { url "https://repository.apache.org/content/groups/snapshots" }
}

Set the Beam dependency versions to the following value: 2.XX.0-SNAPSHOT.

This configuration directs the build system to download Beam nightly builds from the Maven Snapshot Repository. The local build that you edited isn‘t downloaded. You usually don’t need to build all Beam artifacts locally. If you do need to build all Beam artifacts locally, use the following command for all projects ./gradlew -Ppublishing publishToMavenLocal.

The following situations require additional consideration.

If you're using the standard Dataflow runner (not Runner v2), and the worker harness has changed, do the following:

Use the following command to compile dataflowWorkerJar:
```
./gradlew :runners:google-cloud-dataflow-java:worker:shadowJar
```
The jar is located in the build output.

Use the following command to pass pipelineOption:

--dataflowWorkerJar=/.../beam-runners-google-cloud-dataflow-java-legacy-worker-2.XX.0-SNAPSHOT.jar

If you're using Dataflow Runner v2 and sdks/java/harness or its dependencies (like sdks/java/core) have changed, do the following:

Use the following command to build the SDK harness container:

./gradlew :sdks:java:container:java8:docker # java8, java11, java17, etc

docker tag apache/beam_java8_sdk:2.49.0.dev
“us.gcr.io/apache-beam-testing/beam_java11_sdk:2.49.0-custom” # change to your container registry docker push “us.gcr.io/apache-beam-testing/beam_java11_sdk:2.49.0-custom” ```

Run the pipeline with the following options:

--experiments=use_runner_v2 \
--sdkContainerImage="us.gcr.io/apache-beam-testing/beam_java11_sdk:2.49.0-custom"

Python guide

The Beam Python SDK is distributed as a single wheel, which is more straightforward than the Java SDK.

Console setup

These instructions explain how to configure your console (shell) for Python development. In this example, the working directory is set to sdks/python.

Recommended: Install the Python interpreter by using pyenv. Use the following commands:
install prerequisites
curl https://pyenv.run | bash
pyenv install 3.X (a supported Python version; see python_version in project property
Use the following commands to set up and activate the virtual environment:
pyenv virtualenv 3.X ENV_NAME
pyenv activate ENV_NAME
Install the apache_beam package in editable mode: pip install -e .[gcp, test]
For development that uses an SDK container image, do the following:
Install Docker Desktop.
Install Go.
If you're going to submit PRs, use the following command to precommit the hook for Python code changes (nobody likes lint failures!!):

# enable pre-commit
(env) $ pip install pre-commit
(env) $ pre-commit install

# disable pre-commit
(env) $ pre-commit uninstall

Run a unit test

Although the tests can be triggered with a Gradle command, that method sets up a new virtualenv and installs dependencies before each run, which takes minutes. Therefore, it's useful to have a persistent virtualenv.

Unit tests have the filename **_test.py.

To run all tests in a file, use the following command:

pytest -v  apache_beam/io/textio_test.py

To run all tests in a class, use the following command:

pytest -v  apache_beam/io/textio_test.py::TextSourceTest

To run a specific test, use the following command:

pytest -v  apache_beam/io/textio_test.py::TextSourceTest::test_progress

Run an integration test

Integration tests have the filename **_it_test.py.

To run an integration test on the Direct Runner, use the following command:

python -m pytest -o log_cli=True -o log_level=Info \
  apache_beam/ml/inference/pytorch_inference_it_test.py::PyTorchInference \
  --test-pipeline-options='--runner=TestDirectRunner’

If you're preparing a PR, for test-suites to run in PostCommit Python, add tests paths under batchTests in the common.gradle file.

To run an integration test on the Dataflow Runner, follow these steps:

To build the SDK tarball, use the following command:

cd sdks/python
pip install build && python -m build --sdist

The tarball file is generated in the sdks/python/sdist/ directory.

To specify the tarball file, use the --test-pipeline-options parameter. Use the location --sdk_location=dist/apache-beam-2.53.0.dev0.tar.gz. The following example shows the complete command:

python -m pytest -o log_cli=True -o log_level=Info \
apache_beam/ml/inference/pytorch_inference_it_test.py::PyTorchInference \
--test-pipeline-options='--runner=TestDataflowRunner --project=<project>
                         --temp_location=gs://<bucket>/tmp
                         --sdk_location=dist/apache-beam-2.35.0.dev0.tar.gz
                         --region=us-central1’

If you‘re preparing a PR, to include integration tests in the Python PostCommit test suite’s Dataflow task, use the marker @pytest.mark.it_postcommit.

Build containers for modified SDK code

To build containers for modified SDK code, follow these steps.

Run the following command:

./gradlew :sdks:python:container:py39:docker \
-Pdocker-repository-root=<gcr.io/location> -Pdocker-tag=<tag>

Push the containers.
Specify the container location by using the --sdk_container_image option.

The following example shows a complete command:

python -m pytest  -o log_cli=True -o log_level=Info \
  apache_beam/ml/inference/pytorch_inference_it_test.py::PyTorchInference \
  --test-pipeline-options='--runner=TestDataflowRunner --project=<project>
                           --temp_location=gs://<bucket>/tmp
                           --sdk_container_image=us.gcr.io/apache-beam-testing/beam-sdk/beam:dev
                           --region=us-central1’

Specify additional test dependencies

This section provides two options for specifying additional test dependencies.

Use the --requirements_file options. The following example demonstrates how to use the --requirements_file options:

python -m pytest  -o log_cli=True -o log_level=Info \
  apache_beam/ml/inference/pytorch_inference_it_test.py::PyTorchInference \
  --test-pipeline-options='--runner=TestDataflowRunner --project=<project>
                           --temp_location=gs://<bucket>/tmp
                           --sdk_location=us.gcr.io/apache-beam-testing/beam-sdk/beam:dev
                           --region=us-central1
                           –requirements_file=requirements.txt’

If you're using the Dataflow runner, use custom containers. You can use the official Beam SDK container image as a base and then apply your changes.

Run your pipeline with modified beam code

To run your pipeline with modified beam code, follow these steps:

Build the Beam SDK tarball. Under sdks/python, run python -m build --sdist. For more details, see Run an integration test on this page.
Install the Apache Beam Python SDK in your Python virtual environment with the necessary extensions. Use a command similar to the following example: pip install /path/to/apache-beam.tar.gz[gcp].
Initiate your Python script. To run your pipeline, use a command similar to the following example:

python my_pipeline.py --runner=DataflowRunner --sdk_location=/path/to/apache-beam.tar.gz --project=my_project --region=us-central1 --temp_location=gs://my-bucket/temp ...

Tips for using the Dataflow runner:

The Python worker installs the Apache Beam SDK before processing work items. Therefore, you don‘t usually need to provide a custom worker container. If your Google Cloud VM doesn’t have internet access and transient dependencies are changed from the officially released container images, you do need to provide a custom worker container. In this case, see Build containers for modified SDK code on this page.
Installing the Beam Python SDK from source can be slow (3.5 minutes for an1-standard-1 machine). As an alternative, if the host machine uses amd64 architecture, you can build a wheel instead of a tarball by using a command similar to ./gradle :sdks:python:bdistPy311linux (for Python 3.11). To pass the built wheel, use the --sdk_location option. That installation completes in seconds.

Caveat - `save_main_session`

NameError when running DoFn on remote runner
Global imports, functions, and variables in main pipeline module are not serialized by default
Use --save_main_session pipeline option to enable it

Appendix

Directories of snapshot builds

https://repository.apache.org/content/groups/snapshots/org/apache/beam/ Java SDK build (nightly)
https://gcr.io/apache-beam-testing/beam-sdk Beam SDK container build (Java, Python, Go, every 4 hrs)
https://gcr.io/apache-beam-testing/beam_portability Portable runner (Flink, Spark) job server container (nightly)
gs://beam-python-nightly-snapshots Python SDK build (nightly)

Beam code change guide

Repository structure

Code paths

Java

Non-Java SDKs

Gradle quick start

Gradle key concepts

Beam-specific Gradle project configuration

Environment setup

Java development guide

IDE (IntelliJ) setup

Console setup

Run a unit test

Run integration tests

Write integration tests

Run your pipeline with modified beam code

Python guide

Console setup

Run a unit test

Run an integration test

Build containers for modified SDK code

Specify additional test dependencies

Run your pipeline with modified beam code

Caveat - save_main_session

Appendix

Directories of snapshot builds

Caveat - `save_main_session`