This PR changes Airflow installation and build backend to use new
standard Python ways of building Python applications.
We've been trying to do it for quite a while. Airflow tranditionally
has been using complex and convoluted build process based on
setuptools and (extremely) custom setup.py file. It survived
migration to Airflow 2.0 and splitting Airlfow monorepo into
Airflow and Providers, adding pre-installed providers and switching
providers to use flit (and follow build standards).
So far tooling in Python ecosystme had not been able to fuflill our
needs and we refrained to develop our own tooling, but finally with
appearance of Hatch (managed by Python Packaging Authority) and
few recent advancements there we are finally able to swtich to
Python standard ways of managing project dependnecy configuration
and project build setup (with a few customizations).
This PR makes airflow build process follow those standard PEPs:
* Airflow has all build configuration stored in pyproject.toml
following PEP 518 which allows any fronted (`pip`, `poetry`,
`hatch`, `flit`, or whatever other frontend is used to
install required build dependendencies to install Airflow
locally and to build distribution pacakges (sdist/wheel)
* Hatchling backend follows PEP 517 for standard source tree and build
backend implementation that allows to execute the build in a
frontend-independent way
* We store all project metadata in pyprooject.toml - following
PEP 621 where all necessary project metadata components were
defined.
* We plug-in into Hatchling "editable build" hooks following
PEP 660. Hatchling internally builds editable wheel that
is used as ephemeral step and communication between backend
and frontend (and this ephemeral wheel is used to make
editable installation of the projeect - suitable for fast
iteration of code without reinstalling the package.
With Airflow having many provider packages in single source tree
where we want to be able to install and develop airflow and
providers together, this is not a small feat to implement the
case wher editable installation has to behave quite a bit
differently when it comes to packaging and dependencies for
editable install (when you want to edit sources directly) and
installable package (where you want to have separate Airflow
package and provider packages). Fortunately the standardisation
efforts in the Python Packaging community and tooling implementing
it had finally made it possible.
Some of the important ways bow this has been achieved:
* Pyproject.toml is generally managed manually, but the part where
provider dependencies and bundle dependencies are used is
automatically updated by a pre-commit whenever provider
dependencies change.
* We have dedicated (generated) `[devel_provider_*]` extras
that are only installing provider dependencies in editable
mode (not the final provider packages). This allows to install
dependencies of providers individually or in groups in the
editable installation of Airflow, without installing provider
packages (i.e. we can use provider code directly from sources
of editable Airflow installation).
* We have some generated `[devel_*]` bundle extras that bundle
together all or selected provider dependencies for installation
in CI image and local editable virtualenv installation.
* We are utilising custom hatchiling build hooks (PEP 660 standard)
that allow to modify 'standard' wheel package on-the-fly when
the wheel is being prepared by adding preinstalled package
dependencies (which are not needed in editable build) and by
removing all devel extras (that are not needed in the PyPI
distributed wheel package). This allows to solve the conundrum
of having different "editable" and "standard" behaviour while
keeping the same project specification in pyproject.toml.
* We added description of how `Hatch` can be employed as build
frontend in order to manage local virtualenv and install Airflow
in editable way easily - while keeping all properties of the
installed application (including working airflow cli and
package metadata discovery) as well as how to use PEP-standard
ways of bulding wheel and sdist packages.
* We have a custom step (following PEP-standards) to inject
airflow-specific build steps - compiling www assets and
generating git commit hash version to display it in the UI
* We also show how all this makes it possible to make it easy to
manage local virtualenvs and editable installations for Airflow
contributors - without vendor lock-in of the build tools as
by following standard PEPs Airflow can be locally and editably
installed by anyone using any build front-end tools following
the standards - whether you use `pip`, `poetry`, `Hatch`, `flit`
or any other frontent build tools, Airflow local installation
and package building will work the same way for all of them,
where both "editable" and "standard" package prepration is
managed by `hatchling` backend in the same way.
* Previously our extras contained a "." which is not normalized
name for extras - `pip` and other tools replaced it automatically
with `_'. This change updates the extra names to contain
'_' rather than '.' in the name. This should be fully backwards
compatible, users will still be able to use "." but it will be
normalized to "_" in Airflow packages.
* Some of the problematic extras (graphviz, docgen) have been
moved out of the core extras to optional ones. Particularly
graphviz has been difficult to install on MacOS ARM. This
is slightly backwards incompatible, but we should treat it
as bugfix - the only missing feature Airflow will not be
able to handle is to produce DAG output as image (and it
only requires to install graphviz to bring it back). The
difficulty of installing graphviz as required dependency
justifies the slight backwards-incompatible change.
* Additionally, this change organizes the documentation around
the extras and dependencies, explaining the reasoning behind
all the different extras we have.
* As a bonus (and this is what we used to test it all) we are
documenting how to use Hatch frontend to:
* manage multiple Python installations
* manage multiple Pythob virtualenv environments
* build Airflow packages for release management