Skip to content

sukovanej/inseminator

Repository files navigation

Inseminator

codecov Code style: black

(definition from dictionary.com)

a technician who introduces prepared semen dependencies into the genital tract of breeding animals python classes, especially cows and mares pure classes with proper IoC, for artificial insemination well coupled components and clear classes signatures.

Python library for type-based dependency injection. Write code without global state and noisy boilerplate. Inseminator is meant to be used in an entry-point layer of your application and the only thing it requires is properly type-hinted classes dependencies.

Installation

Install using the pip tool.

pip install inseminator

Usage

You start by defining the container of your dependencies. Whenever you want the container to resolve a dependency, it uses the container to search for existing objects and a resolver automatically creates desired dependencies.

from inseminator import Container


class DomainModel:
    def __init__(self):
        self.__logic_constant = 1

    def domain_logic(self, input_value: int) -> int:
        return input_value + self.__logic_constant


class Controller:
    def __init__(self, domain_model: DomainModel):
        self.__domain_model = domain_model

    def handler(self, input_value: int) -> int:
        return self.__domain_model.domain_logic(input_value)


# entry-point of your application

container = Container()

# view layer handling

controller = container.resolve(Controller)
result = controller.handler(1)
print(result)

The strategy for resolving Controller is its constructor signature. The resolver works as follows.

  1. We ask the container to resolve a dependency Controller -> container.resolve(Controller).
  2. Resolver inside the container checks the Controller's constructor signature, i.e. type hints of __init__ method and sees domain_models: DomainModel.
  3. If an instance of DomainModel class is already known by the container it uses that instance. In the opposite case, the container starts the same resolving machinery for DomainModel - which is the exact case we are facing now.
  4. Because DomainModel doesn't have any dependencies it can construct it directly.
  5. Now the resolver has all the dependencies for Controller constructor and can instantiate it.

If we programmed against an interface instead of implementation the example is modified like this.

from inseminator import Container

from typing import Protocol

class DomainModel(Protocol):
    def domain_logic(self, input_value: int) -> int:
        ...

class Controller:
    def __init__(self, domain_model: DomainModel):
        self.__domain_model = domain_model

    def handler(self, input_value: int) -> int:
        return self.__domain_model.domain_logic(input_value)


# domain model implementation


class ConcreteDomainModel:
    def __init__(self):
        self.__logic_constant = 1

    def domain_logic(self, input_value: int) -> int:
        return input_value + self.__logic_constant


# entry point of your application

container = Container()
container.register(DomainModel, value=ConcreateDomainModel())

# view layer handling

controller = container.resolve(Controller)
result = controller.handler(1)
print(result)

In this situation, protocol DomainModel doesn't hold implementation details, only interface. Using

container.register(DomainModel, value=ConcreateDomainModel())

we're guiding the resolver to use instance of ConcreateDomainModel in case someone asks for DomainModel.

Enforced parameters

If it is not desired to provide a single concrete implementation for abstract or protocol dependency one can enforce the resolver to use concrete types for specified parameters. Simply call container.resolve also with keywords and tell the resolve how it should resolve some particular parameters.

container = Container()
controller = container.resolve(Controller, domain_model=ConcreteDomainModel())

Moreover, using this approach ConcreteDomainModel is not evaluated and saved in the container but rather in a sub-container which exists only during the resolving. Therefore, if we want to create another instance that depends on DomainModel we must either use register or again specify the parameter during resolving.

Injecting functions

It might be convinient to specify funcion's dependencies in-place. The great example is Flask handler function. It should live in the same layer the DI container lives because it provides only infrastructure functionality and desirably the only thing it does it calling domain layer's functions. For this purpose, there is injector decorator on the Container object. You just tell which dependency to provide using Depends type constructor.

from inseminator import Container, Depends


class Dependency:
    def __init__(self):
        self.x = 1


container = Container()


@container.inject
def my_handler(input_value: int, dependency: Dependency = Depends(Dependency)):
    return input_value + dependency.x

Used like that, my_handler takes a single argument and thanks to closure it has dependency prepared with the right instance of Dependency.

>>> my_handler(1)
2

Default parameter values

When default parameters are specified the resolver uses them unless we override that value by enforced parameter.

def MyDependency:
    def __init__(self, parameter: int = 1) -> None:
        self.parameter = parameter

my_dependency = container.resolve(MyDependency)
assert my_dependency.parameter == 1