Skip to content

Latest commit

 

History

History
1468 lines (1021 loc) · 67.1 KB

CONTRIBUTING.rst

File metadata and controls

1468 lines (1021 loc) · 67.1 KB
Raw

Contributions

Contributions are welcome and are greatly appreciated! Every little bit helps, and credit will always be given.

This document aims to explain the subject of contributions if you have not contributed to any Open Source project, but it will also help people who have contributed to other projects learn about the rules of that community.

New Contributor

If you are a new contributor, please follow the Contributors Quick Start guide to get a gentle step-by-step introduction to setting up the development environment and making your first contribution.

Get Mentoring Support

If you are new to the project, you might need some help in understanding how the dynamics of the community works and you might need to get some mentorship from other members of the community - mostly committers. Mentoring new members of the community is part of committers job so do not be afraid of asking committers to help you. You can do it via comments in your Pull Request, asking on a devlist or via Slack. For your convenience, we have a dedicated #newbie-questions Slack channel where you can ask any questions you want - it's a safe space where it is expected that people asking questions do not know a lot about Airflow (yet!).

If you look for more structured mentoring experience, you can apply to Apache Software Foundation's Official Mentoring Programme. Feel free to follow it and apply to the programme and follow up with the community.

Report Bugs

Report bugs through GitHub.

Please report relevant information and preferably code that exhibits the problem.

Fix Bugs

Look through the GitHub issues for bugs. Anything is open to whoever wants to implement it.

Issue reporting and resolution process

The Apache Airflow project uses a set of labels for tracking and triaging issues, as well as a set of priorities and milestones to track how and when the enhancements and bug fixes make it into an Airflow release. This is documented as part of the Issue reporting and resolution process,

Implement Features

Look through the GitHub issues labeled "kind:feature" for features.

Any unassigned feature request issue is open to whoever wants to implement it.

We've created the operators, hooks, macros and executors we needed, but we've made sure that this part of Airflow is extensible. New operators, hooks, macros and executors are very welcomed!

Improve Documentation

Airflow could always use better documentation, whether as part of the official Airflow docs, in docstrings, docs/*.rst or even on the web as blog posts or articles.

Submit Feedback

The best way to send feedback is to open an issue on GitHub.

If you are proposing a new feature:

  • Explain in detail how it would work.
  • Keep the scope as narrow as possible to make it easier to implement.
  • Remember that this is a volunteer-driven project, and that contributions are welcome :)

Roles

There are several roles within the Airflow Open-Source community.

For detailed information for each role, see: Committers and PMC's.

PMC Member

The PMC (Project Management Committee) is a group of maintainers that drives changes in the way that Airflow is managed as a project.

Considering Apache, the role of the PMC is primarily to ensure that Airflow conforms to Apache's processes and guidelines.

Committers/Maintainers

Committers are community members that have write access to the project’s repositories, i.e., they can modify the code, documentation, and website by themselves and also accept other contributions.

The official list of committers can be found here.

Additionally, committers are listed in a few other places (some of these may only be visible to existing committers):

Committers are responsible for:

  • Championing one or more items on the Roadmap
  • Reviewing & Merging Pull-Requests
  • Scanning and responding to GitHub issues
  • Responding to questions on the dev mailing list ([email protected])

Contributors

A contributor is anyone who wants to contribute code, documentation, tests, ideas, or anything to the Apache Airflow project.

Contributors are responsible for:

  • Fixing bugs
  • Adding features
  • Championing one or more items on the Roadmap.

Contribution Workflow

Typically, you start your first contribution by reviewing open tickets at GitHub issues.

If you create pull-request, you don't have to create an issue first, but if you want, you can do it. Creating an issue will allow you to collect feedback or share plans with other people.

For example, you want to have the following sample ticket assigned to you: #7782: Add extra CC: to the emails sent by Airflow.

In general, your contribution includes the following stages:

Contribution Workflow

  1. Make your own fork of the Apache Airflow main repository.
  2. Create a local virtualenv, initialize the Breeze environment, and install pre-commit framework. If you want to add more changes in the future, set up your fork and enable GitHub Actions.
  3. Join devlist and set up a Slack account.
  4. Make the change and create a Pull Request from your fork.
  5. Ping @ #development slack, comment @people. Be annoying. Be considerate.

Step 1: Fork the Apache Airflow Repo

From the apache/airflow repo, create a fork:

Creating a fork

Step 2: Configure Your Environment

You can use either a local virtual env or a Docker-based env. The differences between the two are explained here.

The local env's instructions can be found in full in the LOCAL_VIRTUALENV.rst file. The Docker env is here to maintain a consistent and common development environment so that you can replicate CI failures locally and work on solving them locally rather by pushing to CI.

You can configure the Docker-based Breeze development environment as follows:

  1. Install the latest versions of the Docker Community Edition and Docker Compose and add them to the PATH.

  2. Enter Breeze: ./breeze

    Breeze starts with downloading the Airflow CI image from the Docker Hub and installing all required dependencies.

  3. Enter the Docker environment and mount your local sources to make them immediately visible in the environment.

  4. Create a local virtualenv, for example:

mkvirtualenv myenv --python=python3.6
  1. Initialize the created environment:
./breeze initialize-local-virtualenv --python 3.6
  1. Open your IDE (for example, PyCharm) and select the virtualenv you created as the project's default virtualenv in your IDE.

Step 3: Connect with People

For effective collaboration, make sure to join the following Airflow groups:

Step 4: Prepare PR

  1. Update the local sources to address the issue.

    For example, to address this example issue, do the following:

    • Read about email configuration in Airflow.
    • Find the class you should modify. For the example GitHub issue, this is email.py.
    • Find the test class where you should add tests. For the example ticket, this is test_email.py.
    • Make sure your fork's master is synced with Apache Airflow's master before you create a branch. See How to sync your fork for details.
    • Create a local branch for your development. Make sure to use latest apache/master as base for the branch. See How to Rebase PR for some details on setting up the apache remote. Note, some people develop their changes directly in their own master branches - this is OK and you can make PR from your master to apache/master but we recommend to always create a local branch for your development. This allows you to easily compare changes, have several changes that you work on at the same time and many more. If you have apache set as remote then you can make sure that you have latest changes in your master by git pull apache master when you are in the local master branch. If you have conflicts and want to override your locally changed master you can override your local changes with git fetch apache; git reset --hard apache/master.
    • Modify the class and add necessary code and unit tests.
    • Run the unit tests from the IDE or local virtualenv as you see fit.
    • Run the tests in Breeze.
    • Run and fix all the static checks. If you have pre-commits installed, this step is automatically run while you are committing your code. If not, you can do it manually via git add and then pre-commit run.

  2. Rebase your fork, squash commits, and resolve all conflicts. See How to rebase PR if you need help with rebasing your change. Remember to rebase often if your PR takes a lot of time to review/fix. This will make rebase process much easier and less painful and the more often you do it, the more comfortable you will feel doing it.

  3. Re-run static code checks again.

  4. Make sure your commit has a good title and description of the context of your change, enough for the committer reviewing it to understand why you are proposing a change. Make sure to follow other PR guidelines described in pull request guidelines. Create Pull Request! Make yourself ready for the discussion!

  5. Depending on "scope" of your changes, your Pull Request might go through one of few paths after approval. We run some non-standard workflow with high degree of automation that allows us to optimize the usage of queue slots in GitHub Actions. Our automated workflows determine the "scope" of changes in your PR and send it through the right path:

    • In case of a "no-code" change, approval will generate a comment that the PR can be merged and no tests are needed. This is usually when the change modifies some non-documentation related RST files (such as this file). No python tests are run and no CI images are built for such PR. Usually it can be approved and merged few minutes after it is submitted (unless there is a big queue of jobs).
    • In case of change involving python code changes or documentation changes, a subset of full test matrix will be executed. This subset of tests perform relevant tests for single combination of python, backend version and only builds one CI image and one PROD image. Here the scope of tests depends on the scope of your changes:
      • when your change does not change "core" of Airflow (Providers, CLI, WWW, Helm Chart) you will get the comment that PR is likely ok to be merged without running "full matrix" of tests. However decision for that is left to committer who approves your change. The committer might set a "full tests needed" label for your PR and ask you to rebase your request or re-run all jobs. PRs with "full tests needed" run full matrix of tests.
      • when your change changes the "core" of Airflow you will get the comment that PR needs full tests and the "full tests needed" label is set for your PR. Additional check is set that prevents from accidental merging of the request until full matrix of tests succeeds for the PR.

    More details about the PR workflow be found in PULL_REQUEST_WORKFLOW.rst.

Step 5: Pass PR Review

PR Review

Note that committers will use Squash and Merge instead of Rebase and Merge when merging PRs and your commit will be squashed to single commit.

You need to have review of at least one committer (if you are committer yourself, it has to be another committer). Ideally you should have 2 or more committers reviewing the code that touches the core of Airflow.

Pull Request Guidelines

Before you submit a pull request (PR) from your forked repo, check that it meets these guidelines:

  • Include tests, either as doctests, unit tests, or both, to your pull request.

    The airflow repo uses GitHub Actions to run the tests and codecov to track coverage. You can set up both for free on your fork. It will help you make sure you do not break the build with your PR and that you help increase coverage.

  • Follow our project's Coding style and best practices.

    These are things that aren't currently enforced programmatically (either because they are too hard or just not yet done.)

  • Rebase your fork, squash commits, and resolve all conflicts.

  • When merging PRs, wherever possible try to use Squash and Merge instead of Rebase and Merge.

  • Add an Apache License header to all new files.

    If you have pre-commit hooks enabled, they automatically add license headers during commit.

  • If your pull request adds functionality, make sure to update the docs as part of the same PR. Doc string is often sufficient. Make sure to follow the Sphinx compatible standards.

  • Make sure your code fulfills all the static code checks we have in our code. The easiest way to make sure of that is to use pre-commit hooks

  • Run tests locally before opening PR.

  • Make sure the pull request works for Python 3.6 and 3.7.

  • Adhere to guidelines for commit messages described in this article. This makes the lives of those who come after you a lot easier.

Airflow Git Branches

All new development in Airflow happens in the master branch. All PRs should target that branch.

We also have a v2-0-test branch that is used to test 2.0.x series of Airflow and where committers cherry-pick selected commits from the master branch.

Cherry-picking is done with the -x flag.

The v2-0-test branch might be broken at times during testing. Expect force-pushes there so committers should coordinate between themselves on who is working on the v2-0-test branch - usually these are developers with the release manager permissions.

The v2-0-stable branch is rather stable - there are minimum changes coming from approved PRs that passed the tests. This means that the branch is rather, well, "stable".

Once the v2-0-test branch stabilises, the v2-0-stable branch is synchronized with v2-0-test. The v2-0-stable branch is used to release 2.0.x releases.

The general approach is that cherry-picking a commit that has already had a PR and unit tests run against main is done to v2-0-test branch, but PRs from contributors towards 2.0 should target v2-0-stable branch.

The v2-0-test branch and v2-0-stable ones are merged just before the release and that's the time when they converge.

The production images are build in DockerHub from:

  • master branch for development
  • v2-0-test branch for testing 2.0.x release
  • 2.0.*, 2.0.*rc* releases from the v2-0-stable branch when we prepare release candidates and final releases. There are no production images prepared from v2-0-stable branch.

Similar rules apply to 1.10.x releases until June 2021. We have v1-10-test and v1-10-stable branches there.

Development Environments

There are two environments, available on Linux and macOS, that you can use to develop Apache Airflow:

The table below summarizes differences between the two environments:

Property Local virtualenv Breeze environment
Test coverage
  • (-) unit tests only
  • (+) integration and unit tests
Setup
  • (+) automated with breeze cmd
  • (+) automated with breeze cmd
Installation difficulty
  • (-) depends on the OS setup
  • (+) works whenever Docker works
Team synchronization
  • (-) difficult to achieve
  • (+) reproducible within team
Reproducing CI failures
  • (-) not possible in many cases
  • (+) fully reproducible
Ability to update
  • (-) requires manual updates
  • (+) automated update via breeze cmd
Disk space and CPU usage
  • (+) relatively lightweight
  • (-) uses GBs of disk and many CPUs
IDE integration
  • (+) straightforward
  • (-) via remote debugging only

Typically, you are recommended to use both of these environments depending on your needs.

Local virtualenv Development Environment

All details about using and running local virtualenv environment for Airflow can be found in LOCAL_VIRTUALENV.rst.

Benefits:

  • Packages are installed locally. No container environment is required.
  • You can benefit from local debugging within your IDE.
  • With the virtualenv in your IDE, you can benefit from autocompletion and running tests directly from the IDE.

Limitations:

  • You have to maintain your dependencies and local environment consistent with other development environments that you have on your local machine.

  • You cannot run tests that require external components, such as mysql, postgres database, hadoop, mongo, cassandra, redis, etc.

    The tests in Airflow are a mixture of unit and integration tests and some of them require these components to be set up. Local virtualenv supports only real unit tests. Technically, to run integration tests, you can configure and install the dependencies on your own, but it is usually complex. Instead, you are recommended to use Breeze development environment with all required packages pre-installed.

  • You need to make sure that your local environment is consistent with other developer environments. This often leads to a "works for me" syndrome. The Breeze container-based solution provides a reproducible environment that is consistent with other developers.

  • You are STRONGLY encouraged to also install and use pre-commit hooks for your local virtualenv development environment. Pre-commit hooks can speed up your development cycle a lot.

Breeze Development Environment

All details about using and running Airflow Breeze can be found in BREEZE.rst.

The Airflow Breeze solution is intended to ease your local development as "It's a Breeze to develop Airflow".

Benefits:

  • Breeze is a complete environment that includes external components, such as mysql database, hadoop, mongo, cassandra, redis, etc., required by some of Airflow tests. Breeze provides a preconfigured Docker Compose environment where all these services are available and can be used by tests automatically.
  • Breeze environment is almost the same as used in the CI automated builds. So, if the tests run in your Breeze environment, they will work in the CI as well. See CI.rst for details about Airflow CI.

Limitations:

  • Breeze environment takes significant space in your local Docker cache. There are separate environments for different Python and Airflow versions, and each of the images takes around 3GB in total.
  • Though Airflow Breeze setup is automated, it takes time. The Breeze environment uses pre-built images from DockerHub and it takes time to download and extract those images. Building the environment for a particular Python version takes less than 10 minutes.
  • Breeze environment runs in the background taking precious resources, such as disk space and CPU. You can stop the environment manually after you use it or even use a bare environment to decrease resource usage.

NOTE: Breeze CI images are not supposed to be used in production environments. They are optimized for repeatability of tests, maintainability and speed of building rather than production performance. The production images are not yet officially published.

Airflow dependencies

Note

Only pip installation is currently officially supported.

While they are some successes with using other tools like poetry or pip-tools, they do not share the same workflow as pip - especially when it comes to constraint vs. requirements management. Installing via Poetry or pip-tools is not currently supported.

If you wish to install airflow using those tools you should use the constraint files and convert them to appropriate format and workflow that your tool requires.

Extras

There are a number of extras that can be specified when installing Airflow. Those extras can be specified after the usual pip install - for example pip install -e .[ssh]. For development purpose there is a devel extra that installs all development dependencies. There is also devel_ci that installs all dependencies needed in the CI environment.

This is the full list of those extras:

airbyte, all, all_dbs, amazon, apache.atlas, apache.beam, apache.cassandra, apache.druid, apache.hdfs, apache.hive, apache.kylin, apache.livy, apache.pig, apache.pinot, apache.spark, apache.sqoop, apache.webhdfs, async, atlas, aws, azure, cassandra, celery, cgroups, cloudant, cncf.kubernetes, crypto, dask, databricks, datadog, devel, devel_all, devel_ci, devel_hadoop, dingding, discord, doc, docker, druid, elasticsearch, exasol, facebook, ftp, gcp, gcp_api, github_enterprise, google, google_auth, grpc, hashicorp, hdfs, hive, http, imap, jdbc, jenkins, jira, kerberos, kubernetes, ldap, microsoft.azure, microsoft.mssql, microsoft.winrm, mongo, mssql, mysql, neo4j, odbc, openfaas, opsgenie, oracle, pagerduty, papermill, password, pinot, plexus, postgres, presto, qds, qubole, rabbitmq, redis, s3, salesforce, samba, segment, sendgrid, sentry, sftp, singularity, slack, snowflake, spark, sqlite, ssh, statsd, tableau, telegram, trino, vertica, virtualenv, webhdfs, winrm, yandex, zendesk

Provider packages

Airflow 2.0 is split into core and providers. They are delivered as separate packages:

  • apache-airflow - core of Apache Airflow
  • apache-airflow-providers-* - More than 50 provider packages to communicate with external services

In Airflow 1.10 all those providers were installed together within one single package and when you installed airflow locally, from sources, they were also installed. In Airflow 2.0, providers are separated out, and not packaged together with the core, unless you set INSTALL_PROVIDERS_FROM_SOURCES environment variable to true.

In Breeze - which is a development environment, INSTALL_PROVIDERS_FROM_SOURCES variable is set to true, but you can add --skip-installing-airflow-providers-from-sources flag to Breeze to skip installing providers when building the images.

One watch-out - providers are still always installed (or rather available) if you install airflow from sources using -e (or --editable) flag. In such case airflow is read directly from the sources without copying airflow packages to the usual installation location, and since 'providers' folder is in this airflow folder - the providers package is importable.

Some of the packages have cross-dependencies with other providers packages. This typically happens for transfer operators where operators use hooks from the other providers in case they are transferring data between the providers. The list of dependencies is maintained (automatically with pre-commits) in the airflow/providers/dependencies.json. Pre-commits are also used to generate dependencies. The dependency list is automatically used during PyPI packages generation.

Cross-dependencies between provider packages are converted into extras - if you need functionality from the other provider package you can install it adding [extra] after the apache-airflow-providers-PROVIDER for example: pip install apache-airflow-providers-google[amazon] in case you want to use GCP transfer operators from Amazon ECS.

If you add a new dependency between different providers packages, it will be detected automatically during pre-commit phase and pre-commit will fail - and add entry in dependencies.json so that the package extra dependencies are properly added when package is installed.

You can regenerate the whole list of provider dependencies by running this command (you need to have pre-commits installed).

pre-commit run build-providers-dependencies

Here is the list of packages and their extras:

Package Extras
airbyte http
amazon apache.hive,exasol,ftp,google,imap,mongo,mysql,postgres,ssh
apache.beam google
apache.druid apache.hive
apache.hive amazon,microsoft.mssql,mysql,presto,samba,vertica
apache.livy http
dingding http
discord http
google amazon,apache.beam,apache.cassandra,cncf.kubernetes,facebook,microsoft.azure,microsoft.mssql,mysql,oracle,postgres,presto,salesforce,sftp,ssh,trino
hashicorp google
microsoft.azure google,oracle
microsoft.mssql odbc
mysql amazon,presto,trino,vertica
opsgenie http
postgres amazon
salesforce tableau
sftp ssh
slack http
snowflake slack

Developing community managed provider packages

While you can develop your own providers, Apache Airflow has 60+ providers that are managed by the community. They are part of the same repository as Apache Airflow (we use monorepo approach where different parts of the system are developed in the same repository but then they are packaged and released separately). All the community-managed providers are in 'airflow/providers' folder and they are all sub-packages of 'airflow.providers' package. All the providers are available as apache-airflow-providers-<PROVIDER_ID> packages.

The capabilities of the community-managed providers are the same as the third-party ones. When the providers are installed from PyPI, they provide the entry-point containing the metadata as described in the previous chapter. However when they are locally developed, together with Airflow, the mechanism of discovery of the providers is based on provider.yaml file that is placed in the top-folder of the provider. Similarly as in case of the provider.yaml file is compliant with the json-schema specification. Thanks to that mechanism, you can develop community managed providers in a seamless way directly from Airflow sources, without preparing and releasing them as packages. This is achieved by:

  • When Airflow is installed locally in editable mode (pip install -e) the provider packages installed from PyPI are uninstalled and the provider discovery mechanism finds the providers in the Airflow sources by searching for provider.yaml files.
  • When you want to install Airflow from sources you can set INSTALL_PROVIDERS_FROM_SOURCES variable to true and then the providers will not be installed from PyPI packages, but they will be installed from local sources as part of the apache-airflow package, but additionally the provider.yaml files are copied together with the sources, so that capabilities and names of the providers can be discovered. This mode is especially useful when you are developing a new provider, that cannot be installed from PyPI and you want to check if it installs cleanly.

Regardless if you plan to contribute your provider, when you are developing your own, custom providers, you can use the above functionality to make your development easier. You can add your provider as a sub-folder of the airflow.providers package, add the provider.yaml file and install airflow in development mode - then capabilities of your provider will be discovered by airflow and you will see the provider among other providers in airflow providers command output.

Documentation for the community managed providers

When you are developing a community-managed provider, you are supposed to make sure it is well tested and documented. Part of the documentation is provider.yaml file integration information and version information. This information is stripped-out from provider info available at runtime, however it is used to automatically generate documentation for the provider.

If you have pre-commits installed, pre-commit will warn you and let you know what changes need to be done in the provider.yaml file when you add a new Operator, Hooks, Sensor or Transfer. You can also take a look at the other provider.yaml files as examples.

Well documented provider contains those:

  • index.rst with references to packages, API used and example dags
  • configuration reference
  • class documentation generated from PyDoc in the code
  • example dags
  • how-to guides

You can see for example google provider which has very comprehensive documentation:

Part of the documentation are example dags. We are using the example dags for various purposes in providers:

  • showing real examples of how your provider classes (Operators/Sensors/Transfers) can be used
  • snippets of the examples are embedded in the documentation via exampleinclude:: directive
  • examples are executable as system tests

Testing the community managed providers

We have high requirements when it comes to testing the community managed providers. We have to be sure that we have enough coverage and ways to tests for regressions before the community accepts such providers.

  • Unit tests have to be comprehensive and they should tests for possible regressions and edge cases not only "green path"
  • Integration tests where 'local' integration with a component is possible (for example tests with MySQL/Postgres DB/Trino/Kerberos all have integration tests which run with real, dockerized components
  • System Tests which provide end-to-end testing, usually testing together several operators, sensors, transfers connecting to a real external system

You can read more about out approach for tests in TESTING.rst but here are some highlights.

Dependency management

Airflow is not a standard python project. Most of the python projects fall into one of two types - application or library. As described in this StackOverflow question, the decision whether to pin (freeze) dependency versions for a python project depends on the type. For applications, dependencies should be pinned, but for libraries, they should be open.

For application, pinning the dependencies makes it more stable to install in the future - because new (even transitive) dependencies might cause installation to fail. For libraries - the dependencies should be open to allow several different libraries with the same requirements to be installed at the same time.

The problem is that Apache Airflow is a bit of both - application to install and library to be used when you are developing your own operators and DAGs.

This - seemingly unsolvable - puzzle is solved by having pinned constraints files. Those are available as of airflow 1.10.10 and further improved with 1.10.12 (moved to separate orphan branches)

Pinned constraint files

Note

Only pip installation is officially supported.

While they are some successes with using other tools like poetry or pip-tools, they do not share the same workflow as pip - especially when it comes to constraint vs. requirements management. Installing via Poetry or pip-tools is not currently supported.

If you wish to install airflow using those tools you should use the constraint files and convert them to appropriate format and workflow that your tool requires.

By default when you install apache-airflow package - the dependencies are as open as possible while still allowing the apache-airflow package to install. This means that apache-airflow package might fail to install in case a direct or transitive dependency is released that breaks the installation. In such case when installing apache-airflow, you might need to provide additional constraints (for example pip install apache-airflow==1.10.2 Werkzeug<1.0.0)

There are several sets of constraints we keep:

  • 'constraints' - those are constraints generated by matching the current airflow version from sources
    and providers that are installed from PyPI. Those are constraints used by the users who want to install airflow with pip, they are named constraints-<PYTHON_MAJOR_MINOR_VERSION>.txt.
  • "constraints-source-providers" - those are constraints generated by using providers installed from current sources. While adding new providers their dependencies might change, so this set of providers is the current set of the constraints for airflow and providers from the current master sources. Those providers are used by CI system to keep "stable" set of constraints. Thet are named constraints-source-providers-<PYTHON_MAJOR_MINOR_VERSION>.txt
  • "constraints-no-providers" - those are constraints generated from only Apache Airflow, without any providers. If you want to manage airflow separately and then add providers individually, you can use those. Those constraints are named constraints-no-providers-<PYTHON_MAJOR_MINOR_VERSION>.txt.

We also have constraints with "source-providers" but they are used i

The first ones can be used as constraints file when installing Apache Airflow in a repeatable way. It can be done from the sources:

pip install -e . \
  --constraint "https://raw.githubusercontent.com/apache/airflow/constraints-master/constraints-3.6.txt"

or from the PyPI package:

pip install apache-airflow \
  --constraint "https://raw.githubusercontent.com/apache/airflow/constraints-master/constraints-3.6.txt"

This works also with extras - for example:

pip install .[ssh] \
  --constraint "https://raw.githubusercontent.com/apache/airflow/constraints-master/constraints-3.6.txt"

As of apache-airflow 1.10.12 it is also possible to use constraints directly from GitHub using specific tag/hash name. We tag commits working for particular release with constraints-<version> tag. So for example fixed valid constraints 1.10.12 can be used by using constraints-1.10.12 tag:

pip install apache-airflow[ssh]==1.10.12 \
    --constraint "https://raw.githubusercontent.com/apache/airflow/constraints-1.10.12/constraints-3.6.txt"

There are different set of fixed constraint files for different python major/minor versions and you should use the right file for the right python version.

If you want to update just airflow dependencies, without paying attention to providers, you can do it using -no-providers constraint files as well.

pip install . --upgrade \
  --constraint "https://raw.githubusercontent.com/apache/airflow/constraints-master/constraints-no-providers-3.6.txt"

The constraints-<PYTHON_MAJOR_MINOR_VERSION>.txt and constraints-no-providers-<PYTHON_MAJOR_MINOR_VERSION>.txt will be automatically regenerated by CI job every time after the setup.py is updated and pushed if the tests are successful.

Manually generating constraint files

The constraint files are generated automatically by the CI job. Sometimes however it is needed to regenerate them manually (committers only). For example when master build did not succeed for quite some time). This can be done by running this (it utilizes parallel preparation of the constraints):

export CURRENT_PYTHON_MAJOR_MINOR_VERSIONS_AS_STRING="3.6 3.7 3.8"
for python_version in $(echo "${CURRENT_PYTHON_MAJOR_MINOR_VERSIONS_AS_STRING}")
do
  ./breeze build-image --upgrade-to-newer-dependencies --python ${python_version} --build-cache-local
  ./breeze build-image --upgrade-to-newer-dependencies --python ${python_version} --build-cache-local
  ./breeze build-image --upgrade-to-newer-dependencies --python ${python_version} --build-cache-local
done

GENERATE_CONSTRAINTS_MODE="pypi-providers" ./scripts/ci/constraints/ci_generate_all_constraints.sh
GENERATE_CONSTRAINTS_MODE="source-providers" ./scripts/ci/constraints/ci_generate_all_constraints.sh
GENERATE_CONSTRAINTS_MODE="no-providers" ./scripts/ci/constraints/ci_generate_all_constraints.sh

AIRFLOW_SOURCES=$(pwd)

The constraints will be generated in "files/constraints-PYTHON_VERSION/constraints-*.txt files. You need to checkout the right 'constraints-' branch in a separate repository and then you can copy, commit and push the generated files:

cd <AIRFLOW_WITH_CONSTRAINT_MASTER_DIRECTORY>
git pull
cp ${AIRFLOW_SOURCES}/files/constraints-*/constraints*.txt .
git diff
git add .
git commit -m "Your commit message here" --no-verify
git push

Documentation

Documentation for apache-airflow package and other packages that are closely related to it ie. providers packages are in /docs/ directory. For detailed information on documentation development, see: docs/README.rst

For Helm Chart documentation, see: chart/README.md

Static code checks

We check our code quality via static code checks. See STATIC_CODE_CHECKS.rst for details.

Your code must pass all the static code checks in the CI in order to be eligible for Code Review. The easiest way to make sure your code is good before pushing is to use pre-commit checks locally as described in the static code checks documentation.

Coding style and best practices

Most of our coding style rules are enforced programmatically by flake8 and pylint (which are run automatically on every pull request), but there are some rules that are not yet automated and are more Airflow specific or semantic than style

Don't Use Asserts Outside Tests

Our community agreed that to various reasons we do not use assert in production code of Apache Airflow. For details check the relevant mailing list thread.

In other words instead of doing:

assert some_predicate()

you should do:

if not some_predicate():
    handle_the_case()

Database Session Handling

Explicit is better than implicit. If a function accepts a session parameter it should not commit the transaction itself. Session management is up to the caller.

To make this easier there is the create_session helper:

from airflow.utils.session import create_session

def my_call(*args, session):
  ...
  # You MUST not commit the session here.

with create_session() as session:
    my_call(*args, session=session)

If this function is designed to be called by "end-users" (i.e. DAG authors) then using the @provide_session wrapper is okay:

from airflow.utils.session import provide_session

...

@provide_session
def my_method(arg, arg, session=None)
  ...
  # You SHOULD not commit the session here. The wrapper will take care of commit()/rollback() if exception

Don't use time() for duration calculations

If you wish to compute the time difference between two events with in the same process, use time.monotonic(), not time.time() nor timzeone.utcnow().

If you are measuring duration for performance reasons, then time.perf_counter() should be used. (On many platforms, this uses the same underlying clock mechanism as monotonic, but perf_counter is guaranteed to be the highest accuracy clock on the system, monotonic is simply "guaranteed" to not go backwards.)

If you wish to time how long a block of code takes, use Stats.timer() -- either with a metric name, which will be timed and submitted automatically:

from airflow.stats import Stats

...

with Stats.timer("my_timer_metric"):
    ...

or to time but not send a metric:

from airflow.stats import Stats

...

with Stats.timer() as timer:
    ...

log.info("Code took %.3f seconds", timer.duration)

For full docs on timer() check out `airflow/stats.py`_.

If the start_date of a duration calculation needs to be stored in a database, then this has to be done using datetime objects. In all other cases, using datetime for duration calculation MUST be avoided as creating and diffing datetime operations are (comparatively) slow.

Naming Conventions for provider packages

In Airflow 2.0 we standardized and enforced naming for provider packages, modules and classes. those rules (introduced as AIP-21) were not only introduced but enforced using automated checks that verify if the naming conventions are followed. Here is a brief summary of the rules, for detailed discussion you can go to AIP-21 Changes in import paths

The rules are as follows:

  • Provider packages are all placed in 'airflow.providers'
  • Providers are usually direct sub-packages of the 'airflow.providers' package but in some cases they can be further split into sub-packages (for example 'apache' package has 'cassandra', 'druid' ... providers ) out of which several different provider packages are produced (apache.cassandra, apache.druid). This is case when the providers are connected under common umbrella but very loosely coupled on the code level.
  • In some cases the package can have sub-packages but they are all delivered as single provider package (for example 'google' package contains 'ads', 'cloud' etc. sub-packages). This is in case the providers are connected under common umbrella and they are also tightly coupled on the code level.
  • Typical structure of provider package:
    • example_dags -> example DAGs are stored here (used for documentation and System Tests)
    • hooks -> hooks are stored here
    • operators -> operators are stored here
    • sensors -> sensors are stored here
    • secrets -> secret backends are stored here
    • transfers -> transfer operators are stored here
  • Module names do not contain word "hooks", "operators" etc. The right type comes from the package. For example 'hooks.datastore' module contains DataStore hook and 'operators.datastore' contains DataStore operators.
  • Class names contain 'Operator', 'Hook', 'Sensor' - for example DataStoreHook, DataStoreExportOperator
  • Operator name usually follows the convention: <Subject><Action><Entity>Operator (BigQueryExecuteQueryOperator) is a good example
  • Transfer Operators are those that actively push data from one service/provider and send it to another service (might be for the same or another provider). This usually involves two hooks. The convention for those <Source>To<Destination>Operator. They are not named *TransferOperator nor *Transfer.
  • Operators that use external service to perform transfer (for example CloudDataTransferService operators are not placed in "transfers" package and do not have to follow the naming convention for transfer operators.
  • It is often debatable where to put transfer operators but we agreed to the following criteria:
    • We use "maintainability" of the operators as the main criteria - so the transfer operator should be kept at the provider which has highest "interest" in the transfer operator
    • For Cloud Providers or Service providers that usually means that the transfer operators should land at the "target" side of the transfer
  • Secret Backend name follows the convention: <SecretEngine>Backend.
  • Tests are grouped in parallel packages under "tests.providers" top level package. Module name is usually test_<object_to_test>.py,
  • System tests (not yet fully automated but allowing to run e2e testing of particular provider) are named with _system.py suffix.

Test Infrastructure

We support the following types of tests:

  • Unit tests are Python tests launched with pytest. Unit tests are available both in the Breeze environment and local virtualenv.
  • Integration tests are available in the Breeze development environment that is also used for Airflow's CI tests. Integration test are special tests that require additional services running, such as Postgres, Mysql, Kerberos, etc.
  • System tests are automatic tests that use external systems like Google Cloud. These tests are intended for an end-to-end DAG execution.

For details on running different types of Airflow tests, see TESTING.rst.

Metadata Database Updates

When developing features, you may need to persist information to the metadata database. Airflow has Alembic built-in module to handle all schema changes. Alembic must be installed on your development machine before continuing with migration.

# starting at the root of the project
$ pwd
~/airflow
# change to the airflow directory
$ cd airflow
$ alembic revision -m "add new field to db"
   Generating
~/airflow/airflow/migrations/versions/12341123_add_new_field_to_db.py

Node.js Environment Setup

airflow/www/ contains all yarn-managed, front-end assets. Flask-Appbuilder itself comes bundled with jQuery and bootstrap. While they may be phased out over time, these packages are currently not managed with yarn.

Make sure you are using recent versions of node and yarn. No problems have been found with node>=8.11.3 and yarn>=1.19.1.

Installing yarn and its packages

Make sure yarn is available in your environment.

To install yarn on macOS:

  1. Run the following commands (taken from this source):
brew install node
brew install yarn
yarn config set prefix ~/.yarn
  1. Add ~/.yarn/bin to your PATH so that commands you are installing could be used globally.
  2. Set up your .bashrc file and then source ~/.bashrc to reflect the change.
export PATH="$HOME/.yarn/bin:$PATH"
  1. Install third-party libraries defined in package.json by running the following commands within the airflow/www/ directory:
# from the root of the repository, move to where our JS package.json lives
cd airflow/www/
# run yarn install to fetch all the dependencies
yarn install

These commands install the libraries in a new node_modules/ folder within www/.

Should you add or upgrade a node package, run yarn add --dev <package> for packages needed in development or yarn add <package> for packages used by the code. Then push the newly generated package.json and yarn.lock file so that we could get a reproducible build. See the Yarn docs for more details.

Generate Bundled Files with yarn

To parse and generate bundled files for Airflow, run either of the following commands:

# Compiles the production / optimized js & css
yarn run prod

# Starts a web server that manages and updates your assets as you modify them
yarn run dev

Follow JavaScript Style Guide

We try to enforce a more consistent style and follow the JS community guidelines.

Once you add or modify any JavaScript code in the project, please make sure it follows the guidelines defined in Airbnb JavaScript Style Guide.

Apache Airflow uses ESLint as a tool for identifying and reporting on patterns in JavaScript. To use it, run any of the following commands:

# Check JS code in .js and .html files, and report any errors/warnings
yarn run lint

# Check JS code in .js and .html files, report any errors/warnings and fix them if possible
yarn run lint:fix

How to sync your fork

When you have your fork, you should periodically synchronize the master of your fork with the Apache Airflow master. In order to do that you can git pull --rebase to your local git repository from apache remote and push the master (often with --force to your fork). There is also an easy way using Force sync master from apache/airflow workflow. You can go to "Actions" in your repository and choose the workflow and manually trigger the workflow using "Run workflow" command.

This will force-push the master from apache/airflow to the master in your fork. Note that in case you modified the master in your fork, you might loose those changes.

How to rebase PR

A lot of people are unfamiliar with the rebase workflow in Git, but we think it is an excellent workflow, providing a better alternative to the merge workflow. We've therefore written a short guide for those who would like to learn it.

As opposed to the merge workflow, the rebase workflow allows us to clearly separate your changes from the changes of others. It puts the responsibility of rebasing on the author of the change. It also produces a "single-line" series of commits on the master branch. This makes it easier to understand what was going on and to find reasons for problems (it is especially useful for "bisecting" when looking for a commit that introduced some bugs).

First of all, we suggest you read about the rebase workflow here: Merging vs. rebasing. This is an excellent article that describes all the ins/outs of the rebase workflow. I recommend keeping it for future reference.

The goal of rebasing your PR on top of apache/master is to "transplant" your change on top of the latest changes that are merged by others. It also allows you to fix all the conflicts that arise as a result of other people changing the same files as you and merging the changes to apache/master.

Here is how rebase looks in practice (you can find a summary below these detailed steps):

1. You first need to add the Apache project remote to your git repository. This is only necessary once, so if it's not the first time you are following this tutorial you can skip this step. In this example, we will be adding the remote as "apache" so you can refer to it easily:

  • If you use ssh: git remote add apache [email protected]:apache/airflow.git
  • If you use https: git remote add apache https://github.com/apache/airflow.git

  1. You then need to make sure that you have the latest master fetched from the apache repository. You can do this via:

    git fetch apache (to fetch apache remote)

    git fetch --all (to fetch all remotes)

  2. Assuming that your feature is in a branch in your repository called my-branch you can easily check what is the base commit you should rebase from by:

    git merge-base my-branch apache/master

    This will print the HASH of the base commit which you should use to rebase your feature from. For example: 5abce471e0690c6b8d06ca25685b0845c5fd270f. Copy that HASH and go to the next step.

    Optionally, if you want better control you can also find this commit hash manually.

    Run:

    git log

    And find the first commit that you DO NOT want to "transplant".

    Performing:

    git rebase HASH

    Will "transplant" all commits after the commit with the HASH.

  3. Providing that you weren't already working on your branch, check out your feature branch locally via:

    git checkout my-branch

  4. Rebase:

    git rebase HASH --onto apache/master

    For example:

    git rebase 5abce471e0690c6b8d06ca25685b0845c5fd270f --onto apache/master

  5. If you have no conflicts - that's cool. You rebased. You can now run git push --force-with-lease to push your changes to your repository. That should trigger the build in our CI if you have a Pull Request (PR) opened already.

  6. While rebasing you might have conflicts. Read carefully what git tells you when it prints information about the conflicts. You need to solve the conflicts manually. This is sometimes the most difficult part and requires deliberately correcting your code and looking at what has changed since you developed your changes.

    There are various tools that can help you with this. You can use:

    git mergetool

    You can configure different merge tools with it. You can also use IntelliJ/PyCharm's excellent merge tool. When you open a project in PyCharm which has conflicts, you can go to VCS > Git > Resolve Conflicts and there you have a very intuitive and helpful merge tool. For more information, see Resolve conflicts.

  7. After you've solved your conflict run:

    git rebase --continue

    And go either to point 6. or 7, depending on whether you have more commits that cause conflicts in your PR (rebasing applies each commit from your PR one-by-one).

Summary

Useful when you understand the flow but don't remember the steps and want a quick reference.

git fetch --all git merge-base my-branch apache/master git checkout my-branch git rebase HASH --onto apache/master git push --force-with-lease

How to communicate

Apache Airflow is a Community within Apache Software Foundation. As the motto of the Apache Software Foundation states "Community over Code" - people in the community are far more important than their contribution.

This means that communication plays a big role in it, and this chapter is all about it.

In our communication, everyone is expected to follow the ASF Code of Conduct.

We have various channels of communication - starting from the official devlist, comments in the Pull Requests, Slack, wiki.

All those channels can be used for different purposes. You can join the channels via links at the Airflow Community page

  • The Apache Airflow devlist for:
    • official communication
    • general issues, asking community for opinion
    • discussing proposals
    • voting
  • The Airflow CWiki for:
    • detailed discussions on big proposals (Airflow Improvement Proposals also name AIPs)
    • helpful, shared resources (for example Apache Airflow logos
    • information that can be re-used by others (for example instructions on preparing workshops)
  • GitHub Pull Requests (PRs) for:
    • discussing implementation details of PRs
    • not for architectural discussions (use the devlist for that)
  • The deprecated JIRA issues for:
    • checking out old but still valuable issues that are not on GitHub yet
    • mentioning the JIRA issue number in the title of the related PR you would like to open on GitHub

IMPORTANT We don't create new issues on JIRA anymore. The reason we still look at JIRA issues is that there are valuable tickets inside of it. However, each new PR should be created on GitHub issues as stated in Contribution Workflow Example

  • The Apache Airflow Slack for:
    • ad-hoc questions related to development (#development channel)
    • asking for review (#development channel)
    • asking for help with PRs (#how-to-pr channel)
    • troubleshooting (#troubleshooting channel)
    • group talks (including SIG - special interest groups) (#sig-* channels)
    • notifications (#announcements channel)
    • random queries (#random channel)
    • regional announcements (#users-* channels)
    • newbie questions (#newbie-questions channel)
    • occasional discussions (wherever appropriate including group and 1-1 discussions)

The devlist is the most important and official communication channel. Often at Apache project you can hear "if it is not in the devlist - it did not happen". If you discuss and agree with someone from the community on something important for the community (including if it is with committer or PMC member) the discussion must be captured and reshared on devlist in order to give other members of the community to participate in it.

We are using certain prefixes for email subjects for different purposes. Start your email with one of those:
  • [DISCUSS] - if you want to discuss something but you have no concrete proposal yet
  • [PROPOSAL] - if usually after "[DISCUSS]" thread discussion you want to propose something and see what other members of the community think about it.
  • [AIP-NN] - if the mail is about one of the Airflow Improvement Proposals
  • [VOTE] - if you would like to start voting on a proposal discussed before in a "[PROPOSAL]" thread

Voting is governed by the rules described in Voting

We are all devoting our time for community as individuals who except for being active in Apache Airflow have families, daily jobs, right for vacation. Sometimes we are in different time zones or simply are busy with day-to-day duties that our response time might be delayed. For us it's crucial to remember to respect each other in the project with no formal structure. There are no managers, departments, most of us is autonomous in our opinions, decisions. All of it makes Apache Airflow community a great space for open discussion and mutual respect for various opinions.

Disagreements are expected, discussions might include strong opinions and contradicting statements. Sometimes you might get two committers asking you to do things differently. This all happened in the past and will continue to happen. As a community we have some mechanisms to facilitate discussion and come to a consensus, conclusions or we end up voting to make important decisions. It is important that these decisions are not treated as personal wins or looses. At the end it's the community that we all care about and what's good for community, should be accepted even if you have a different opinion. There is a nice motto that you should follow in case you disagree with community decision "Disagree but engage". Even if you do not agree with a community decision, you should follow it and embrace (but you are free to express your opinion that you don't agree with it).

As a community - we have high requirements for code quality. This is mainly because we are a distributed and loosely organised team. We have both - contributors that commit one commit only, and people who add more commits. It happens that some people assume informal "stewardship" over parts of code for some time - but at any time we should make sure that the code can be taken over by others, without excessive communication. Setting high requirements for the code (fairly strict code review, static code checks, requirements of automated tests, pre-commit checks) is the best way to achieve that - by only accepting good quality code. Thanks to full test coverage we can make sure that we will be able to work with the code in the future. So do not be surprised if you are asked to add more tests or make the code cleaner - this is for the sake of maintainability.

Here are a few rules that are important to keep in mind when you enter our community:

  • Do not be afraid to ask questions
  • The communication is asynchronous - do not expect immediate answers, ping others on slack (#development channel) if blocked
  • There is a #newbie-questions channel in slack as a safe place to ask questions
  • You can ask one of the committers to be a mentor for you, committers can guide within the community
  • You can apply to more structured Apache Mentoring Programme
  • It’s your responsibility as an author to take your PR from start-to-end including leading communication in the PR
  • It’s your responsibility as an author to ping committers to review your PR - be mildly annoying sometimes, it’s OK to be slightly annoying with your change - it is also a sign for committers that you care
  • Be considerate to the high code quality/test coverage requirements for Apache Airflow
  • If in doubt - ask the community for their opinion or propose to vote at the devlist
  • Discussions should concern subject matters - judge or criticise the merit but never criticise people
  • It’s OK to express your own emotions while communicating - it helps other people to understand you
  • Be considerate for feelings of others. Tell about how you feel not what you think of others

Commit Policy

The following commit policy passed by a vote 8(binding FOR) to 0 against on May 27, 2016 on the dev list and slightly modified and consensus reached in October 2020:

  • Commits need a +1 vote from a committer who is not the author
  • Do not merge a PR that regresses linting or does not pass CI tests (unless we have justification such as clearly transient error).
  • When we do AIP voting, both PMC and committer +1s are considered as binding vote.

Resources & Links