The outline for this document in GitHub is available at top-right corner button (with 3-dots and 3 lines).
This document gathers the necessary steps to create a new community provider and also guidelines for updating the existing ones. You should be aware that providers may have distinctions that may not be covered in this guide. The sequence described was designed to meet the most linear flow possible in order to develop a new provider.
Another recommendation that will help you is to look for a provider that works similar to yours. That way it will help you to set up tests and other dependencies.
First, you need to set up your local development environment. See
Contributors Quick Start
if you did not set up your local environment yet. We recommend using breeze to develop locally. This way you
easily be able to have an environment more similar to the one executed by GitHub CI workflow.
./breeze
Using the code above you will set up Docker containers. These containers your local code to internal volumes. In this way, the changes made in your IDE are already applied to the code inside the container and tests can be carried out quickly.
In this how-to guide our example provider name will be <PROVIDER>.
When you see this placeholder you must change for your provider name.
Most likely you have developed a version of the provider using some local customization and now you need to transfer this code to the Airflow project. Below is described all the initial code structure that the provider may need. Understand that not all providers will need all the components described in this structure. If you still have doubts about building your provider, we recommend that you read the initial provider guide and open a issue on GitHub so the community can help you.
The folders are optional: example_dags, hooks, links, logs, notifications, operators, secrets, sensors, transfers, triggers (and the list changes continuously).
GIT apache/airflow/ └── providers/ └── <PROVIDER>/ ├── pyproject.toml ├── provider.yaml ├── src/ │ └── airflow/ │ └── providers/<PROVIDER>/ │ ├── __init__.py │ ├── executors/ │ │ ├── __init__.py │ │ └── *.py │ ├── hooks/ │ │ ├── __init__.py │ │ └── *.py │ ├── notifications/ │ │ ├── __init__.py │ │ └── *.py │ ├── operators/ │ │ ├── __init__.py │ │ └── *.py │ ├── transfers/ │ │ ├── __init__.py │ │ └── *.py │ └── triggers/ │ ├── __init__.py │ └── *.py └── tests/ ├── unit\ │ └── <PROVIDER>/ │ ├── __init__.py │ ├── executors/ │ │ ├── __init__.py │ │ └── test_*.py │ ├── hooks/ │ │ ├── __init__.py │ │ └── test_*.py │ ├── notifications/ │ │ ├── __init__.py │ │ └── test_*.py │ ├── operators/ │ │ ├── __init__.py │ │ └── test_*.py │ ├── transfers/ │ │ ├── __init__.py │ │ └── test_*.py │ └── triggers/ │ ├── __init__.py │ └── test_*.py ├── integration/<PROVIDER>/ │ ├── __init__.py │ └── test_integration_*.py └── system/<PROVIDER>/ ├── __init__.py └── example_*.py
Considering that you have already transferred your provider's code to the above structure, it will now be necessary to create unit tests for each component you created. The example below I have already set up an environment using breeze and I'll run unit tests for my Hook.
root@fafd8d630e46:/opt/airflow# python -m pytest providers/<PROVIDER>/tests/<PROVIDER>/hook/test_*.py
Sometimes we want to release provider that depends on the version of airflow that has not yet been released
- for example when we released common.io provider it had apache-airflow>=2.8.0 dependency.
Providers that have "min-airflow-version" set to the new, upcoming versions should be excluded in
all previous versions of compatibility check matrix in PROVIDERS_COMPATIBILITY_TESTS_MATRIX in
src/airflow_breeze/global_constants.py. Please add it to all previous versions
This is controlled by chicken_egg_providers property in Selective Checks - and our CI will automatically
build and use those chicken-egg providers during the CI process if pre-release version of Airflow is built.
The short provider id (common.io for example) for such a provider should be added
to CHICKEN_EGG_PROVIDERS list in src/airflow_breeze/utils/selective_checks.py:
This list will be kept here until the official version of Airflow the chicken-egg-providers depend on
is released and the version of airflow is updated in the main and v2-X-Y branch to 2.X+1.0.dev0
and 2.X.1.dev0 respectively. After that the chicken-egg providers will be correctly installed because
both 2.X.1.dev0 and 2.X+1.0.dev0 are considered by pip as >2.X.0 (unlike 2.X.0.dev0).
The release process for Airflow includes cleaning the list after Airflow release is published, so the provider will be removed from the list by release manager.
The problem when generating constraints with chicken-eggo providers and building docker image for
pre-release versions of Airflow - because pip does not recognize the .dev0 or .b1
suffixes of those packages as valid in the >=X.Y.Z comparison.
When you want to install a provider package with apache-airflow>=2.8.0 requirement and you have
2.9.0.dev0 airflow package, pip will not install the package, because it does not recognize
2.9.0.dev0 as a valid version for >=2.8.0 dependency. This is because pip
currently implements the minimum version selection algorithm requirement specified in packaging as
described in the packaging version specification
https://packaging.python.org/en/latest/specifications/version-specifiers/#handling-of-pre-releases
Currently pip only allows to include pre-release versions for all installed packages using --pre
flag, but it does not have the possibility of selectively using this flag to only one package.
In order to implement our desired behaviour, we need the case where only apache-airflow is considered
as pre-release version while all the other dependencies only have stable versions and this is currently
not possible.
To work around this limitation, we have introduced the concept of "chicken-egg" providers. Those providers
are providers that are released together with the version of Airflow they depend on. They are released
with the same version number as the Airflow version they depend on, but with a different suffix. For example
apache-airflow-providers-common-io==2.9.0.dev0 is a chicken-egg provider for apache-airflow==2.9.0.dev0.
However - we should not release providers with such exclusion to pypi, so in order to allow our
CI to work with pre-release versions and perform both - constraint generation and image releasing,
we introduced workarounds in our tooling where in case we build a pre-release version of Airflow,
we will locally build the chicken-egg providers from sources and they are installed from local
directory instead of from PyPI.
This workaround might be removed if pip implements the possibility of selectively using --pre flag
for only one package (Which is foreseen as a possibility in the packaging specification but not implemented
by pip).
Note
The current solution of building pre-release images will not work well if the chicken-egg-provider is
pre-installed package because slim imges will not use the chicken-egg-provider. This could be solved
by adding --chicken-egg-providers flag to slim image building step in released_dockerhub_image.yml
but it would also require filtering out the non-pre-installed packages from it, so the current solution
is to assume pre-installed packages are not chicken-egg providers.
An important part of building a new provider is the documentation.
Some steps for documentation occurs automatically by pre-commit see
Installing pre-commit guide
Those are important files in the airflow source tree that affect providers. The pyproject.toml in root
Airflow folder is automatically generated based on content of provider.yaml file in each provider
when pre-commit is run. Files such as extra-packages-ref.rst should be manually updated because
they are manually formatted for better layout and pre-commit will just verify if the information
about provider is updated there. Files like commit.rst and CHANGELOG are automatically updated
by breeze release-management command by release manager when providers are released.
├── pyproject.toml └── providers/<PROVIDER>/src/airflow/providers/ ├── provider.yaml ├── pyproject.toml ├── CHANGELOG.rst │ └── docs/ ├── integration-logos │ └── <PROVIDER>.png ├── index.rst ├── commits.rst ├── connections.rst └── operators/ └── <PROVIDER>.rst
There is a chance that your provider's name is not a common English word.
In this case is necessary to add it to the file docs/spelling_wordlist.txt.
Add your provider dependencies into provider.yaml under dependencies key..
If your provider doesn't have any dependency add a empty list.
In the docs/apache-airflow-providers-<PROVIDER>/connections.rst:
- add information how to configure connection for your provider.
In the provider's docs/operators/<PROVIDER>.rst add information
how to use the Operator. It's important to add examples and additional information if your
Operator has extra-parameters.
.. _howto/operator:NewProviderOperator: NewProviderOperator =================== Use the :class:`~airflow.providers.<PROVIDER>.operators.NewProviderOperator` to do something amazing with Airflow! Using the Operator ^^^^^^^^^^^^^^^^^^ The NewProviderOperator requires a ``connection_id`` and this other awesome parameter. You can see an example below: .. exampleinclude:: /../../providers/<PROVIDER>/example_dags/example_<PROVIDER>.py :language: python :start-after: [START howto_operator_<PROVIDER>] :end-before: [END howto_operator_<PROVIDER>]
Copy from another, similar provider the docs: docs/*.rst:
At least those docs should be present
- security.rst
- changelog.rst
- commits.rst
- index.rst
- installing-providers-from-sources.rst
- configurations-ref.rst - if your provider has
configelement in provider.yaml with configuration options specific for your provider
Make sure to update/add all information that are specific for the new provider.
In the providers/<PROVIDER>/src/airflow/providers/<PROVIDER>/provider.yaml add information of your provider:
package-name: apache-airflow-providers-<PROVIDER> name: <PROVIDER> description: | `<PROVIDER> <https://example.io/>`__ versions: - 1.0.0 integrations: - integration-name: <PROVIDER> external-doc-url: https://www.example.io/ logo: /docs/integration-logos/<PROVIDER>.png how-to-guide: - /docs/apache-airflow-providers-<PROVIDER>/operators/<PROVIDER>.rst tags: [service] operators: - integration-name: <PROVIDER> python-modules: - airflow.providers.<PROVIDER>.operators.<PROVIDER> hooks: - integration-name: <PROVIDER> python-modules: - airflow.providers.<PROVIDER>.hooks.<PROVIDER> sensors: - integration-name: <PROVIDER> python-modules: - airflow.providers.<PROVIDER>.sensors.<PROVIDER> connection-types: - hook-class-name: airflow.providers.<PROVIDER>.hooks.<PROVIDER>.NewProviderHook - connection-type: provider-connection-type
After changing and creating these files you can build the documentation locally. The two commands below will serve to accomplish this. The first will build your provider's documentation. The second will ensure that the main Airflow documentation that involves some steps with the providers is also working.
breeze build-docs <provider id> breeze build-docs apache-airflow
Those steps above are usually enough for most providers that are "standalone" and not imported or used by other providers (in most cases we will not suspend such providers). However some extra steps might be needed for providers that are used by other providers, or that are part of the default PROD Dockerfile:
- Most of the tests for the suspended provider, will be automatically excluded by pytest collection. However,
in case a provider is dependent on by another provider, the relevant tests might fail to be collected or
run by
pytest. In such cases you should skip the whole test module failing to be collected by addingpytest.importorskipat the top of the test module. For example if your tests fail because they need to importapache.airflow.providers.googleand you have suspended it, you should add this line at the top of the test module that fails.
Example failing collection after google provider has been suspended:
_____ ERROR collecting providers/apache/beam/tests/apache/beam/operators/test_beam.py ______ ImportError while importing test module '/opt/airflow/providers/apache/beam/tests/apache/beam/operators/test_beam.py'. Hint: make sure your test modules/packages have valid Python names. Traceback: /usr/local/lib/python3.8/importlib/__init__.py:127: in import_module return _bootstrap._gcd_import(name[level:], package, level) providers/apache/beam/tests/apache/beam/operators/test_beam.py:25: in <module> from airflow.providers.apache.beam.operators.beam import ( airflow/providers/apache/beam/operators/beam.py:35: in <module> from airflow.providers.google.cloud.hooks.dataflow import ( airflow/providers/google/cloud/hooks/dataflow.py:32: in <module> from google.cloud.dataflow_v1beta3 import GetJobRequest, Job, JobState, JobsV1Beta3AsyncClient, JobView E ModuleNotFoundError: No module named 'google.cloud.dataflow_v1beta3' _ ERROR collecting providers/microsoft/azure/tests/microsoft/azure/transfers/test_azure_blob_to_gcs.py _
The fix is to add this line at the top of the providers/apache/beam/tests/apache/beam/operators/test_beam.py module:
pytest.importorskip("apache.airflow.providers.google")
Some of the other providers might also just import unconditionally the suspended provider and they will fail during the provider verification step in CI. In this case you should turn the provider imports into conditional imports. For example when import fails after
amazonprovider has been suspended:Traceback (most recent call last): File "/opt/airflow/scripts/in_container/verify_providers.py", line 266, in import_all_classes _module = importlib.import_module(modinfo.name) File "/usr/local/lib/python3.8/importlib/__init__.py", line 127, in import_module return _bootstrap._gcd_import(name, package, level) File "<frozen importlib._bootstrap>", line 1006, in _gcd_import File "<frozen importlib._bootstrap>", line 983, in _find_and_load File "<frozen importlib._bootstrap>", line 967, in _find_and_load_unlocked File "<frozen importlib._bootstrap>", line 677, in _load_unlocked File "<frozen importlib._bootstrap_external>", line 728, in exec_module File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed File "/usr/local/lib/python3.8/site-packages/airflow/providers/mysql/transfers/s3_to_mysql.py", line 23, in <module> from airflow.providers.amazon.aws.hooks.s3 import S3Hook ModuleNotFoundError: No module named 'airflow.providers.amazon'
or:
Error: The airflow.providers.microsoft.azure.transfers.azure_blob_to_gcs object in transfers list in
airflow/providers/microsoft/azure/provider.yaml does not exist or is not a module:
No module named 'gcloud.aio.storage'
The fix for that is to turn the feature into an optional provider feature (in the place where the excluded
airflow.providers import happens:
try: from airflow.providers.amazon.aws.hooks.s3 import S3Hook except ImportError as e: from airflow.exceptions import AirflowOptionalProviderFeatureException raise AirflowOptionalProviderFeatureException(e)
- In case we suspend an important provider, which is part of the default Dockerfile you might want to
update the tests for PROD docker image in
docker_tests/test_prod_image.py. - Some of the suspended providers might also fail
breezeunit tests that expect a fixed set of providers. Those tests should be adjusted (but this is not very likely to happen, because the tests are using only the most common providers that we will not be likely to suspend).
We regularly bump min airflow version for all providers we release. This bump is done according to our Provider policies and it is only applied to non-suspended/removed providers. We are running basic import compatibility checks in our CI and the compatibility checks should be updated when min airflow version is updated.
Details on how this should be done are described in Provider policies
Sometimes providers need to have different variants for different versions of Airflow. This is done by:
- copying
version_compat.pyfrom one of the providers that already have conditional variants to the root package of the provider you are working on - importing the
AIRFLOW_V_X_Y_PLUSthat you need from that importedversion_compat.pyfile.
The main reasons we are doing it in this way:
- checking version >= in Python has a non-obvious problem that the pre-release version is always considered
lower than the final version. This is why we are using
AIRFLOW_V_X_Y_PLUSto check for the version that is greater or equal to the version we are checking against - because we want the RC candidates to be considered as equal to the final version (because those RC candidates already contain the feature that is added in the final version). - We do not want to add dependencies to another provider (say
common.compat) without strong need - Even if the code is duplicated, it is just one
version_compat.pyfile that is wholly copied and it is not a big deal to maintain it. - There is a potential risk of one provider importing the same
AIRFLOW_V_X_Y_PLUSfrom another provider (and introduce accidental dependency) or from test code (which should not happen), but we are preventing it via pre-commit checkcheck-imports-in-providersthat will fail if theversion_compatmodule is imported from another provider or from test code.
When releasing providers for the first time, you need to release them in state not-ready.
This will make it available for release management commands, but it will not be added to airflow's
preinstalled providers list - allowing airflow in main CI builds to be built without expecting the
provider to be available in PyPI.
You need to add --include-not-ready-providers if you want to add them to the list of providers
considered by the release management commands.
As soon as the provider is released, you should update the provider to state: ready.
Sometimes we might want to release provider for previous MAJOR when new release is already
released (or bumped in main). This is done by releasing them from providers-<PROVIDER>/vX-Y branch
- for example providers-fab/v1-5 can be used to release the 1.5.2 when 2.0.0 is already being
released or voted on.
The release process looks like usual, the only difference is that the specific branch is used to release the provider and update all documentation, the changes and cherry-picking should be targeting that branch.
As of April 2023, we have the possibility to suspend individual providers, so that they are not holding back dependencies for Airflow and other providers. The process of suspending providers is described in description of the process
Technically, suspending a provider is done by setting state: suspended, in the provider.yaml of the
provider. This should be followed by committing the change and either automatically or manually running
pre-commit checks that will either update derived configuration files or ask you to update them manually.
Note that you might need to run pre-commit several times until all the static checks pass,
because modification from one pre-commit might impact other pre-commits.
If you have pre-commit installed, pre-commit will be run automatically on commit. If you want to run it
manually after commit, you can run it via breeze static-checks --last-commit some of the tests might fail
because suspension of the provider might cause changes in the dependencies, so if you see errors about
missing dependencies imports, non-usable classes etc., you will need to build the CI image locally
via breeze build-image --python 3.9 --upgrade-to-newer-dependencies after the first pre-commit run
and then run the static checks again.
If you want to be absolutely sure to run all static checks you can always do this via
pre-commit run --all-files or breeze static-checks --all-files.
Some of the manual modifications you will have to do (in both cases pre-commit will guide you on what
to do.
- You will have to run
breeze setup regenerate-command-imagesto regenerate breeze help files - you will need to update
extra-packages-ref.rstand in some cases - when mentioned there explicitly -pyproject.tomlto remove the provider from list of dependencies.
What happens under-the-hood as a result, is that pyproject.toml file is updated with
the information about available providers and their dependencies and it is used by our tooling to
exclude suspended providers from all relevant parts of the build and CI system (such as building CI image
with dependencies, building documentation, running tests, etc.)
Resuming providers is done by reverting the original change that suspended it. In case there are changes needed to fix problems in the reverted provider, our CI will detect them and you will have to fix them as part of the PR reverting the suspension.
When removing providers from Airflow code, we need to make one last release where we mark the provider as removed - in documentation and in description of the PyPI package. In order to that release manager has to add "state: removed" flag in the provider yaml file and include the provider in the next wave of the providers (and then remove all the code and documentation related to the provider).
The "removed: removed" flag will cause the provider to be available for the following commands (note that such provider has to be explicitly added as selected to the package - such provider will not be included in the available list of providers or when documentation is built unless --include-removed-providers flag is used):
breeze build-docsbreeze release-management prepare-provider-documentationbreeze release-management prepare-provider-packagesbreeze release-management publish-docs
For all those commands, release manager needs to specify --include-removed-providers when all providers
are built or must add the provider id explicitly during the release process.
Except the changelog that needs to be maintained manually, all other documentation (main page of the provider
documentation, PyPI README), will be automatically updated to include removal notice.