Continuous Integration is important component of making Apache Airflow robust and stable. We are running a lot of tests for every pull request, for main and v2-*-test branches and regularly as CRON jobs.
Our execution environment for CI is GitHub Actions. GitHub Actions (GA) are very well integrated with GitHub code and Workflow and it has evolved fast in 2019/202 to become a fully-fledged CI environment, easy to use and develop for, so we decided to switch to it. Our previous CI system was Travis CI.
However part of the philosophy we have is that we are not tightly coupled with any of the CI environments we use. Most of our CI jobs are written as bash scripts which are executed as steps in the CI jobs. And we have a number of variables determine build behaviour.
You can also take a look at the CI Sequence Diagrams for more graphical overview of how Airlfow's CI works.
Our builds on CI are highly optimized. They utilise some of the latest features provided by GitHub Actions environment that make it possible to reuse parts of the build process across different Jobs.
Big part of our CI runs use Container Images. Airflow has a lot of dependencies and in order to make sure that we are running tests in a well configured and repeatable environment, most of the tests, documentation building, and some more sophisticated static checks are run inside a docker container environment. This environment consist of two types of images: CI images and PROD images. CI Images are used for most of the tests and checks where PROD images are used in the Kubernetes tests.
In order to run the tests, we need to make sure that the images are built using latest sources and that it is done quickly (full rebuild of such image from scratch might take ~15 minutes). Therefore optimisation techniques have been implemented that use efficiently cache from the GitHub Docker registry - in most cases this brings down the time needed to rebuild the image to ~4 minutes. In some cases (when dependencies change) it can be ~6-7 minutes and in case base image of Python releases new patch-level, it can be ~12 minutes.
We are using GitHub Container Registry to store the results of the Build Images
workflow which is used in the Tests workflow.
Currently in main version of Airflow we run tests in 4 different versions of Python (3.7, 3.8, 3.9, 3.10)
which means that we have to build 8 images (4 CI ones and 4 PROD ones). Yet we run around 12 jobs
with each of the CI images. That is a lot of time to just build the environment to run. Therefore
we are utilising pull_request_target feature of GitHub Actions.
This feature allows to run a separate, independent workflow, when the main workflow is run -
this separate workflow is different than the main one, because by default it runs using main version
of the sources but also - and most of all - that it has WRITE access to the GitHub Container Image registry.
This is especially important in our case where Pull Requests to Airflow might come from any repository, and it would be a huge security issue if anyone from outside could utilise the WRITE access to the Container Image Registry via external Pull Request.
Thanks to the WRITE access and fact that the pull_request_target by default uses the main version of the
sources, we can safely run some logic there will checkout the incoming Pull Request, build the container
image from the sources from the incoming PR and push such image to an GitHub Docker Registry - so that
this image can be built only once and used by all the jobs running tests. The image is tagged with unique
COMMIT_SHA of the incoming Pull Request and the tests run in the Pull Request can simply pull such image
rather than build it from the scratch. Pulling such image takes ~ 1 minute, thanks to that we are saving
a lot of precious time for jobs.
We use GitHub Container Registry.
GITHUB_TOKEN is needed to push to the registry and we configured scopes of the tokens in our jobs
to be able to write to the registry.
The main goal of the CI philosophy we have that no matter how complex the test and integration infrastructure, as a developer you should be able to reproduce and re-run any of the failed checks locally. One part of it are pre-commit checks, that allow you to run the same static checks in CI and locally, but another part is the CI environment which is replicated locally with Breeze.
You can read more about Breeze in BREEZE.rst but in essence it is a script that allows
you to re-create CI environment in your local development instance and interact with it. In its basic
form, when you do development you can run all the same tests that will be run in CI - but locally,
before you submit them as PR. Another use case where Breeze is useful is when tests fail on CI. You can
take the full COMMIT_SHA of the failed build pass it as --image-tag parameter of Breeze and it will
download the very same version of image that was used in CI and run it locally. This way, you can very
easily reproduce any failed test that happens in CI - even if you do not check out the sources
connected with the run.
You can read more about it in BREEZE.rst and TESTING.rst
Depending whether the scripts are run locally (most often via Breeze) or whether they
are run in Build Images or Tests workflows they can take different values.
You can use those variables when you try to reproduce the build locally.
The scripts and configuration files for CI jobs are all in scripts/ci - so that in the
pull_request_target target workflow, we can copy those scripts from the main branch and use them
regardless of the changes done in the PR. This way we are kept safe from PRs injecting code into the builds.
build_airflow- builds airflow packagesconstraints- scripts to build and publish latest set of valid constraintsdocs- scripts to build documentationimages- scripts to build and push CI and PROD imageskubernetes- scripts to setup kubernetes cluster, deploy airflow and run kubernetes tests with itopenapi- scripts to run openapi generationpre_commit- scripts to run pre-commit checksprovider_packages- scripts to build and test provider packagesstatic_checks- scripts to run static checks manuallytesting- scripts that run unit and integration teststools- scripts that can be used for various clean-up and preparation tasks
Common libraries of functions for all the scripts can be found in libraries folder. The dockerfiles,
mysql.d, openldap, spectral_rules folders contains DockerFiles and configuration of integrations
needed to run tests.
For detailed use of those scripts you can refer to .github/workflows/ - those scripts are used
by the CI workflows of ours. There are some variables that you can set to change the behaviour of the
scripts.
The default values are "sane" you can change them to interact with your own repositories or registries. Note that you need to set "CI" variable to true in order to get the same results as in CI.
| Variable | Default | Comment |
|---|---|---|
| CI | false |
If set to "true", we simulate behaviour of all scripts as if they are in CI environment |
| CI_TARGET_REPO | apache/airflow |
Target repository for the CI job. Used to compare incoming changes from PR with the target. |
| CI_TARGET_BRANCH | main |
Target branch where the PR should land. Used to compare incoming changes from PR with the target. |
| CI_BUILD_ID | 0 |
Unique id of the build that is kept across re runs
(for GitHub actions it is GITHUB_RUN_ID) |
| CI_JOB_ID | 0 |
Unique id of the job - used to produce unique artifact names. |
| CI_EVENT_TYPE | pull_request |
Type of the event. It can be one of
[
|
| CI_REF | refs/head/main |
Branch in the source repository that is used to make the pull request. |
Our CI uses GitHub Registry to pull and push images to/from by default. You can use your own repo by changing
GITHUB_REPOSITORY and providing your own GitHub Username and Token.
| Variable | Default | Comment |
|---|---|---|
| GITHUB_REPOSITORY | apache/airflow |
Prefix of the image. It indicates which. registry from GitHub to use |
| GITHUB_USERNAME | Username to use to login to GitHub | |
| GITHUB_TOKEN | Token to use to login to GitHub. Only used when pushing images on CI. | |
| GITHUB_REGISTRY_PULL_IMAGE_TAG | latest |
Pull this image tag. This is "latest" by default, can also be full-length commit SHA. |
| GITHUB_REGISTRY_PUSH_IMAGE_TAG | latest |
Push this image tag. This is "latest" by default, can also be full-length commit SHA. |
We are using GitHub Container Registry as cache for our images. Authentication uses GITHUB_TOKEN mechanism. Authentication is needed for pushing the images (WRITE) only in "push", "pull_request_target" workflows.
The following components are part of the CI infrastructure
- Apache Airflow Code Repository - our code repository at https://github.com/apache/airflow
- Apache Airflow Forks - forks of the Apache Airflow Code Repository from which contributors make Pull Requests
- GitHub Actions - (GA) UI + execution engine for our jobs
- GA CRON trigger - GitHub Actions CRON triggering our jobs
- GA Workers - virtual machines running our jobs at GitHub Actions (max 20 in parallel)
- GitHub Image Registry - image registry used as build cache for CI jobs. It is at https://ghcr.io/apache/airflow
- DockerHub Image Registry - image registry used to pull base Python images and (manually) publish the released Production Airflow images. It is at https://dockerhub.com/apache/airflow
- Official Images (future) - these are official images that are prominently visible in DockerHub.
We aim our images to become official images so that you will be able to pull them
with
docker pull apache-airflow
The following CI Job run types are currently run for Apache Airflow (run by ci.yaml workflow) and each of the run types has different purpose and context.
Those runs are results of PR from the forks made by contributors. Most builds for Apache Airflow fall into this category. They are executed in the context of the "Fork", not main Airflow Code Repository which means that they have only "read" permission to all the GitHub resources (container registry, code repository). This is necessary as the code in those PRs (including CI job definition) might be modified by people who are not committers for the Apache Airflow Code Repository.
The main purpose of those jobs is to check if PR builds cleanly, if the test run properly and if the PR is ready to review and merge. The runs are using cached images from the Private GitHub registry - CI, Production Images as well as base Python images that are also cached in the Private GitHub registry. Also for those builds we only execute Python tests if important files changed (so for example if it is "no-code" change, no tests will be executed.
Those runs are results of direct pushes done by the committers or as result of merge of a Pull Request by the committers. Those runs execute in the context of the Apache Airflow Code Repository and have also write permission for GitHub resources (container registry, code repository). The main purpose for the run is to check if the code after merge still holds all the assertions - like whether it still builds, all tests are green.
This is needed because some of the conflicting changes from multiple PRs might cause build and test failures after merge even if they do not fail in isolation. Also those runs are already reviewed and confirmed by the committers so they can be used to do some housekeeping: - pushing most recent image build in the PR to the GitHub Container Registry (for caching) - upgrading to latest constraints and pushing those constraints if all tests succeed - refresh latest Python base images in case new patch-level is released
The housekeeping is important - Python base images are refreshed with varying frequency (once every few months usually but sometimes several times per week) with the latest security and bug fixes. Those patch level images releases can occasionally break Airflow builds (specifically Docker image builds based on those images) therefore in PRs we only use latest "good" Python image that we store in the GitHub Container Registry and those push requests will refresh the latest images if they changed.
Those runs are results of (nightly) triggered job - only for main branch. The
main purpose of the job is to check if there was no impact of external dependency changes on the Apache
Airflow code (for example transitive dependencies released that fail the build). It also checks if the
Docker images can be built from the scratch (again - to see if some dependencies have not changed - for
example downloaded package releases etc.
All runs consist of the same jobs, but the jobs behave slightly differently or they are skipped in different run categories. Here is a summary of the run categories with regards of the jobs they are running. Those jobs often have matrix run strategy which runs several different variations of the jobs (with different Backend type / Python version, type of the tests to run for example). The following chapter describes the workflows that execute for each run.
Those runs and their corresponding Build Images runs are only executed in main apache/airflow
repository, they are not executed in forks - we want to be nice to the contributors and not use their
free build minutes on GitHub Actions.
A general note about cancelling duplicated workflows: for the Build Images, Tests and CodeQL
workflows we use the concurrency feature of GitHub actions to automatically cancel "old" workflow runs
of each type -- meaning if you push a new commit to a branch or to a pull request and there is a workflow
running, GitHub Actions will cancel the old workflow run automatically.
This workflow builds images for the CI Workflow for Pull Requests coming from forks.
It's a special type of workflow: pull_request_target which means that it is triggered when a pull request
is opened. This also means that the workflow has Write permission to push to the GitHub registry the images
used by CI jobs which means that the images can be built only once and reused by all the CI jobs
(including the matrix jobs). We've implemented it so that the Tests workflow waits
until the images are built by the Build Images workflow before running.
Those "Build Image" steps are skipped in case Pull Requests do not come from "forks" (i.e. those are internal PRs for Apache Airflow repository. This is because in case of PRs coming from Apache Airflow (only committers can create those) the "pull_request" workflows have enough permission to push images to GitHub Registry.
This workflow is not triggered on normal pushes to our "main" branches, i.e. after a
pull request is merged and whenever scheduled run is triggered. Again in this case the "CI" workflow
has enough permissions to push the images. In this case we simply do not run this workflow.
The workflow has the following jobs:
| Job | Description |
|---|---|
| Build Info | Prints detailed information about the build |
| Build CI images | Builds all configured CI images |
| Build PROD images | Builds all configured PROD images |
The images are stored in the GitHub Container Registry and the names of those images follow the patterns described in Naming conventions for stored images
Image building is configured in "fail-fast" mode. When any of the images
fails to build, it cancels other builds and the source Tests workflow run
that triggered it.
This workflow is a regular workflow that performs all checks of Airflow code.
| Job | Description | PR | Push Merge | CRON (1) |
|---|---|---|---|---|
| Build info | Prints detailed information about the build | Yes | Yes | Yes |
| Push early cache & images | Pushes early cache/images to GitHub Registry | Yes | Yes | |
| Test OpenAPI client gen | Tests if OpenAPIClient continues to generate | Yes | Yes | Yes |
| UI tests | React UI tests for new Airflow UI | Yes | Yes | Yes |
| WWW tests | React tests for current Airflow UI | Yes | Yes | Yes |
| Test image building | Tests if PROD image build examples work | Yes | Yes | Yes |
| CI Images | Waits for and verify CI Images (3) | Yes | Yes | Yes |
| (Basic) Static checks | Performs static checks (full or basic) | Yes | Yes | Yes |
| Build docs | Builds documentation | Yes | Yes | Yes |
| Tests | Run all the Pytest tests for Python code | Yes(2) | Yes | Yes |
| Tests provider packages | Tests if provider packages work | Yes | Yes | Yes |
| Upload coverage | Uploads test coverage from all the tests | Yes | ||
| PROD Images | Waits for and verify PROD Images (3) | Yes | Yes | Yes |
| Tests Kubernetes | Run Kubernetes test | Yes(2) | Yes | Yes |
| Constraints | Upgrade constraints to latest ones (4) | Yes | Yes | |
| Push cache & images | Pushes cache/images to GitHub Registry (4) | Yes | Yes |
Comments:
- CRON jobs builds images from scratch - to test if everything works properly for clean builds
- The tests are run when the Trigger Tests job determine that important files change (this allows for example "no-code" changes to build much faster)
- The jobs wait for CI images to be available.
- PROD and CI cache & images are pushed as "latest" to GitHub Container registry and constraints are upgrade only if all tests are successful. The images are rebuilt in this step using constraints pushed in the previous step.
The CodeQL security scan uses GitHub security scan framework to scan our code for security violations. It is run for JavaScript and Python code.
Documentation from the main branch is automatically published on Amazon S3.
To make this possible, GitHub Action has secrets set up with credentials
for an Amazon Web Service account - DOCS_AWS_ACCESS_KEY_ID and DOCS_AWS_SECRET_ACCESS_KEY.
This account has permission to write/list/put objects to bucket apache-airflow-docs. This bucket has
public access configured, which means it is accessible through the website endpoint.
For more information, see:
Hosting a static website on Amazon S3
Website endpoint: http://apache-airflow-docs.s3-website.eu-central-1.amazonaws.com/
The images produced during the Build Images workflow of CI jobs are stored in the
GitHub Container Registry
The images are stored with both "latest" tag (for last main push image that passes all the tests as well with the COMMIT_SHA id for images that were used in particular build.
The image names follow the patterns (except the Python image, all the images are stored in
https://ghcr.io/ in apache organization.
The packages are available under (CONTAINER_NAME is url-encoded name of the image). Note that "/" are
supported now in the ghcr.io as apart of the image name within apache organization, but they
have to be percent-encoded when you access them via UI (/ = %2F)
https://github.com/apache/airflow/pkgs/container/<CONTAINER_NAME>
| Image | Name:tag (both cases latest version and per-build) | Description |
|---|---|---|
| Python image (DockerHub) | python:<X.Y>-slim-bullseye | Base Python image used by both production and CI image. Python maintainer release new versions of those image with security fixes every few weeks in DockerHub. |
| Airflow python base image | airflow/<BRANCH>/python:<X.Y>-slim-bullseye | Version of python base image used in Airflow Builds We keep the "latest" version only to mark last "good" python base that went through testing and was pushed. |
| PROD Build image | airflow/<BRANCH>/prod-build/python<X.Y>:latest | Production Build image - this is the "build" stage of production image. It contains build-essentials and all necessary apt packages to build/install PIP packages. We keep the "latest" version only to speed up builds. |
| Manifest CI image | airflow/<BRANCH>/ci-manifest/python<X.Y>:latest | CI manifest image - this is the image used to optimize pulls and builds for Breeze development environment They store hash indicating whether the image will be faster to build or pull. We keep the "latest" version only to help breeze to check if new image should be pulled. |
| CI image | airflow/<BRANCH>/ci/python<X.Y>:latest or airflow/<BRANCH>/ci/python<X.Y>:<COMMIT_SHA> | CI image - this is the image used for most of the tests. Contains all provider dependencies and tools useful For testing. This image is used in Breeze. |
| PROD image | airflow/<BRANCH>/prod/python<X.Y>:latest or airflow/<BRANCH>/prod/python<X.Y>:<COMMIT_SHA> | faster to build or pull. Production image. This is the actual production image optimized for size. It contains only compiled libraries and minimal set of dependencies to run Airflow. |
- <BRANCH> might be either "main" or "v2-*-test"
- <X.Y> - Python version (Major + Minor).Should be one of ["3.7", "3.8", "3.9"].
- <COMMIT_SHA> - full-length SHA of commit either from the tip of the branch (for pushes/schedule) or commit from the tip of the branch used for the PR.
Since we store images from every CI run, you should be able easily reproduce any of the CI tests problems
locally. You can do it by pulling and using the right image and running it with the right docker command,
For example knowing that the CI job was for commit cd27124534b46c9688a1d89e75fcd137ab5137e3:
docker pull ghcr.io/apache/airflow/main/ci/python3.7:cd27124534b46c9688a1d89e75fcd137ab5137e3
docker run -it ghcr.io/apache/airflow/main/ci/python3.7:cd27124534b46c9688a1d89e75fcd137ab5137e3But you usually need to pass more variables and complex setup if you want to connect to a database or enable some integrations. Therefore it is easiest to use Breeze for that. For example if you need to reproduce a MySQL environment in python 3.8 environment you can run:
breeze --image-tag cd27124534b46c9688a1d89e75fcd137ab5137e3 --python 3.8 --backend mysqlYou will be dropped into a shell with the exact version that was used during the CI run and you will be able to run pytest tests manually, easily reproducing the environment that was used in CI. Note that in this case, you do not need to checkout the sources that were used for that run - they are already part of the image - but remember that any changes you make in those sources are lost when you leave the image as the sources are not mapped from your host machine.
In the main branch of development line we currently support Python 3.7, 3.8, 3.9, 3.10
In order to add a new version the following operations should be done (example uses Python 3.10)
- copy the latest constraints in
constraints-mainbranch from previous versions and name it using the new Python version (constraints-3.10.txt). Commit and push - build image locally for both prod and CI locally using Breeze:
breeze ci-image build --python 3.10- Find the 2 new images (prod, ci) created in
GitHub Container registry
go to Package Settings and turn on
Public Visibilityand set "Inherit access from Repository" flag.