Antidote is a dependency injection micro-framework for Python 3.7+, featuring:
- Strong focus on typing and putting type hints to work
- Scalable from small/simple usage to very right "framework frameworks"
It is built on the idea of having a declarative, explicit and decentralized definition of dependencies at the type / function / variable definition. These definitions can be easily tracked down, including by static tooling and startup-time analysis.
Features are built with a strong focus on maintainability, simplicity and ease of use in mind. Everything is statically typed (mypy & pyright), documented with tested examples, and can be easily used in existing code and tested in isolation.
To install Antidote, simply run this command:
pip install antidote
Feel free to open an issue or a discussion on Github for questions, issues, proposals, etc. !
Tutorial, reference and more can be found in the documentation. Some quick links:
Antidote works with a Catalog
which is a sort of "collection" of dependencies. Multiple collections can co-exist, but world
is used by default. The most common form of a dependency is an instance of a given class:
from antidote import injectable
@injectable
class Service:
pass
world[Service] # retrieve the instance
world.get(Service, default='something') # similar to a dict
By default, @injectable
defines a singleton. However, alternative lifetimes (how long the world
keeps value alive in its cache) can exist, such as transient
, where nothing is cached at all.
Dependencies can also be injected into a function/method with @inject
. For both kinds of callables, Mypy, Pyright and PyCharm will infer the correct types.
from antidote import inject
@inject # ⯆ Infers the dependency from the type hint
def f(service: Service = inject.me()) -> Service:
return service
f() # service injected
f(Service()) # useful for testing: no injection, argument is used
@inject
supports a variety of ways to bind arguments to their dependencies (if any.) This binding is always explicit:
from antidote import InjectMe
# recommended with inject.me() for best static-typing experience
@inject
def f2(service = inject[Service]):
...
@inject(kwargs={'service': Service})
def f3(service):
...
@inject
def f4(service: InjectMe[Service]):
...
Classes can also be fully wired, with all methods injected, by using @wire
. It is also possible to
inject the first argument, commonly named self
, of a method with an instance of a class:
@injectable
class Dummy:
@inject.method
def method(self) -> 'Dummy':
return self
# behaves like a class method
assert Dummy.method() is world[Dummy]
# useful for testing: when accessed trough an instance, no injection
dummy = Dummy()
assert dummy.method() is dummy
Antidote comes out-of-the-box with 4 kinds of dependencies:
@injectable
classes for which an instance is provided.from antidote import injectable # ⯆ optional: would just call Service() otherwise. @injectable(factory_method='load') class Service: @classmethod def load(cls) -> 'Service': return cls() world[Service]
const
for defining simple constants.from antidote import const # Used as namespace class Conf: TMP_DIR = const('/tmp') # From environment variables, lazily retrieved LOCATION = const.env("PWD") USER = const.env() # uses the name of the variable PORT = const.env(convert=int) # convert the environment variable to a given type UNKNOWN = const.env(default='unknown') world[Conf.TMP_DIR] @inject def f(tmp_dir: str = inject[Conf.TMP_DIR]): ...
@lazy
function calls (taking into account arguments) used for (stateful-)factories, parameterized dependencies, complex constants, etc.from dataclasses import dataclass from antidote import lazy @dataclass class Template: name: str # the wrapped template function is only executed when accessed through world/@inject @lazy def template(name: str) -> Template: return Template(name=name) # By default a singleton, so it always returns the same instance of Template world[template(name="main")] @inject def f(main_template: Template = inject[template(name="main")]): ...
@lazy
will automatically apply@inject
and can also be a value, property or even a method similarly to@inject.method
.@interface
for which one or multiple implementations can be provided.from antidote import interface, implements @interface class Task: pass @implements(Task) class CustomTask(Task): pass world[Task] # instance of CustomTask
The interface does not need to be a class. It can also be a
Protocol
, a function or a@lazy
function call!@interface def callback(event: str) -> bool: ... @implements(callback) def on_event(event: str) -> bool: # do stuff return True # returns the on_event function assert world[callback] is on_event
@implements
will enforce as much as possible that the interface is correctly implemented. Multiple implementations can also be retrieved. Conditions, filters on metadata and weighting implementation are all supported to allow full customization of which implementation should be retrieved in which use case.
Each of those have several knobs to adapt them to your needs which are covered in the documentation.
Injected functions can typically be tested by passing arguments explicitly but it's not always enough. Antidote provides a test context for full test isolation. The test context allows overriding any dependencies:
original = world[Service]
with world.test.clone() as overrides:
# dependency value is different, but it's still a singleton Service instance
assert world[Service] is not original
# override examples
overrides[Service] = 'x'
assert world[Service] == 'x'
del overrides[Service]
assert world.get(Service) is None
@overrides.factory(Service)
def build_service() -> object:
return 'z'
# Test context can be nested and it wouldn't impact the current test context
with world.test.clone() as nested_overrides:
...
# Outside the test context, nothing changed.
assert world[Service] is original
Antidote also provides introspection capabilities with world.debug
which returns a nicely-formatted tree to show what Antidote actually sees, without actually executing anything:
🟉 <lazy> f()
└── ∅ Service
└── Service.__init__
└── 🟉 <const> Conf.HOST
∅ = transient
↻ = bound
🟉 = singleton
- Scopes are supported. Defining a
ScopeGlobalVar
and using it as a dependency will force any dependents to be updated whenever it changes (a request for example). - Multiple catalogs can be used which lets you expose only a subset of your API (dependencies) to your consumer within a catalog.
- You can easily define your kind of dependencies with proper typing from both
world
andinject
.@injectable
,@lazy
,inject.me()
etc.. all rely on Antidote's core (Provider
,Dependency
, etc.) which is part of the public API.
Check out the Guide which goes more in depth or the Reference for specific features.
- Check for open issues or open a fresh issue to start a discussion around a feature or a bug.
- Fork the repo on GitHub. Run the tests to confirm they all pass on your machine. If you cannot find why it fails, open an issue.
- Start making your changes to the master branch.
- Send a pull request.
Be sure to merge the latest from "upstream" before making a pull request!
If you have any issue during development or just want some feedback, don't hesitate to open a pull request and ask for help ! You're also more than welcome to open a discussion or an issue on any topic!
But, no code changes will be merged if they do not pass mypy, pyright, don't have 100% test coverage or documentation with tested examples (if relevant.)