typedi

typedi

Simple yet powerful typed annotation-based dependency injection container.

To install from python package index simply type (no dependencies):

pip install typedi

Dependency Injection

Dependency Injection (DI) - is a design pattern where an object automatically receives other objects that it depends on. It is an approach to implement Inversion of Control - way to separate concerns and reduce coupling.

To learn more I suggest reading Inversion of Control Containers and the Dependency Injection pattern by Martin Fowler.

When DI in python is useful (my subjective opinion):

• Large modularized applications
• Frameworks foundation
• Plugin systems

When it is harmful or might bring unnecessary complexity:

• Small applications
• Simple elegant pythonic projects
• Large untyped, poorly structured code (will require lots of refactoring in a first place)

typedi

typedi - implements "constructor based" dependency injection by heavily utilizing python's type annotations. This means that you specify object dependencies just by type annotations in __init__ methods or by annotating dataclasses.

From looking at object initializers typedi figures out dependencies and tries its best to provide them:

class AbstractService:
    pass


class ConcreteService(AbstractService):
    pass


class App:
    # App depends on AbstractService.
    # Such that we can provide multiple implementations, i.e. for testing purposes. 
    def __init__(self, service: AbstractService):  
        self.service = service


from typedi import Container
container = Container()  # creating a DI container

# Registering instance of a ConcreteService into container.
container.register_instance(ConcreteService())

# Registering App class.
# Container will inspect the signature of __init__.  
container.register_class(App)  

# First, container tries to specify dependencies of App:
#   * The container looks up registered providers for anything that matches AbstractService.
#   * There is an instance of ConcreteService which matches AbstractService, so it is a match.
# Then instance of App is created with a ConcreteService instance value for `service` argument. 
application = container.resolve(App)

This example is unrealistically simple. But it highlights the fundamental idea of specifying and resolving dependencies from type signatures.

Here is more complex example highlighting some more features:

from typing import Protocol, List, runtime_checkable
from dataclasses import dataclass


# Note: we are using protocol
# Abstract classes (abc) or basic inheritance would also work
@runtime_checkable
class MathOperation(Protocol):
    def apply(self, a: int, b: int) -> int:
        ...


class SumOperation:   # Note: not even subclassing
    def apply(self, a: int, b: int) -> int:
        return a + b


class SubtractOperation:   # Note: not even subclassing
    def apply(self, a: int, b: int) -> int:
        return a - b


@dataclass  # dataclasses (or even attrs) are supported
class App:
    # Require instances that support MathOperation Protocol     
    operations: List[MathOperation]

    # Note: typedi also supports dependencies specified using complex types:
    #   List, Iterable, Union, Optional, Tuple, Type and even Any

    def do_something(self, a: int, b: int):
        for op in self.operations:
            print(f"{op}: {op.apply(a, b)}")


# Importing here just to highlight that it is not needed anywhere above
from typedi import Container

container = Container()

# Register operations as singletons. 
# There will be at most 1 instance in the container 
# Instances will be created only when requested.
container.register_singleton_class(SumOperation)
container.register_singleton_class(SubtractOperation)

# Registering App so that container will understand its dependencies.
container.register_class(App)

# Resolution happens here.
app = container.resolve(App)

app.do_something(3, 4)

typedi aims 0 user code invasion - your code is yours:

• no magic meta subclassing
• no decorators
• no magic Depends default arguments
• no magic __ attributes

You could use typedi in existing application without altering existing class/factory code.

Container

The central concept of most DI implementations including typedi is Container. It serves two purposes:

• Stores registered dependency providers
• Provides interface to query constructed instances

To create container:

from typedi import Container

container = Container()

Registering providers

To be able to resolve dependencies Container should know how to provide them, what instances to provide, what functions to call or what classes to instantiate.

typedi supports multiple different ways of registering instance providers.

• container.register_instance(SomeClass(...)) - Container will just remember this instance and will provide it as a dependency for any object that requires type SomeClass. This could be considered as container-scoped singletons that you manually created.
• container.register_factory(fn) - instead of providing complete initialized instance you could use a factory function that creates it. Container will look at the function signature and remember that instances of return type could be created by calling fn. Every time a return type is requested container will call fn to create an instance.
• container.register_class(SomeClass) - same as container.register_factory but for classes. Container will remember that objects of SomeClass or any parent type in class hierarchy could be created by instantiating SomeClass.
• container.register_singleton_factory(fn) - same as container.register_factory but will be only called once per container and result will be cached.
• container.register_singleton_class(fn) - same as container.register_class but will be only called once per container and result will be cached.

Note: it is possible to register multiple providers of the same type.

Resolving instances

Once container knows how to provide instances you can query them. Objects are queried from the container by some queried type.

• container.resolve(T) - will return an single instance of type T.
• container.get_all_instances(T) - will return a list of instances of type T.
• container.iter_all_instances(T) - will return an iterable of instances of type T.

In each case container will try to find providers that can provide requested type and call them if needed and return resulting instances.

If container is not able to provide an instance(s) of given type it raises ResolutionError.

Supported providers

Here is the list of things recognised by the Container. This means that you can register a factory or a class that produces type -> T and container will be able to resolve it.

Signature	Description
A	some class as a factory of its instances
A(...)	any already created instance
fn(...) -> A	factory of a single instance
fn(...) -> Optional[T]	factory of and optional instance of type T
fn(...) -> Union[T1,...,Tn]	factory of either types
fn(...) -> List[T]	factory of many instances of type T
fn(...) -> Iterable[T]	factory of many instances of T, might be a be a generator as well.
fn(...) -> Tuple[T1, T2]	factory of tuple of T1 and T2 as well as T1 and T1 instances separately.
fn(...) -> Any	factory of any instances (types will be runtime checked).
fn(...) -> Type[T]	factory of generic meta type of T.

• A is any concrete class type.
• T - any supported type including concrete classes, or generics Optional[T], Union[T1, T2], Iterable[T], List[T] and Tuple[T1, ..., TN]. Which means you can go crazy and return something like -> Iterable[Optional[List[Union[A, B]]]]. Container will deal with it.
• ... - is valid annotated signature with 0 or more arguments with definitions, for example:
  • arg: T - argument of some specific type T. If container is not able to resolve T it will raise ResolutionError.
  • arg: Optional[T] - if container will fail to resolve T, None value will be used.
  • *args: T - varargs will be resolved as Tuple[T]. If it is not possible to resolve T - empty tuple() will be passed.
  • arg: T = default_value - container will try to resolve T but if it fails, default_value will be used.

Container will perform a .resolve() for each argument by its annotation. For more information about resolution see below.

Few examples of valid factory/class signatures that you can register:

A()
f() -> A
f(b: B) -> A
f(*args: B) -> A
f(bs: List[B]) -> A
f(b: Optional[B]) -> A
f(bs: Iterable[B]) -> A
f(b: Optional[B] = None) -> A
f(*args: B, arg: C = DEFAULT_C) -> A
f() -> Tuple[List[A], B, C]
f() -> Union[A, B, C]
f() -> Tuple[A, B, C]
f() -> Iterable[A]
f(*args: A) -> Any
f(a: A) -> A
A(a: A)

These could also be used as factories:

• functlools.partial of any typed callable
• Custom callable classes (with annotated __call__)
• @dataclass, @attrs

Not supported return annotations:

• Final
• Literal
• Annotated
• generic annotations with TypeVar

Supported resolution type queries

Query (T)	Description
A	Resolves an instance of A, raises ResolutionError if not possible
Optional[T]	Resolves an instance of type T if possible, otherwise returns None
Union[T1,...,Tn]	Tries resolution of each type argument left to right, if nothing resolves raises ResolutionError
List[T]	Resolves all instances of T and returns a list. If none of T availble, provides empty list.
Iterable[T]	Resolves all instances of T into an iterable. If non of T available, provides empty iterable.
Tuple[T1,...,Tn]	First, tries to resolve an instance of Tuple[T1,...,Tn] as is. If it is not available provides instances of T1 to Tn as a tuple.

typedi also resolves covariant types and even covariant generics like List, Tuple, Iterable and even higher order Type[T].

Covariance example:

from typing import List


class Base:
    pass


class Concrete(Base):
    pass


from typedi import Container
container = Container()

# Registering concrete instances
container.register_instance(Concrete())
container.register_instance(Concrete())

# Query all instances that are instance of Base
concrete_instances = container.resolve(List[Base])

assert isinstance(concrete_instances[0], Concrete)
assert isinstance(concrete_instances[1], Concrete)

Features

typedi also has support of various features:

• Factory functions (including functools.partial)
• Singletons support, both for classes and factory functions
• Optionals support - ability to implement "try resolve or return None if no dependency" behavior

If you want to learn more, please refer to typedi_tests and actual implementation since it is quite self-describing :)

Circular dependencies

typedi can handle circular dependencies solving "chicken and egg" problem:

class A:
    def __init__(self, b: "B"):
        self.b = b

class B:
    def __init__(self, a: A):
        self.a = a

from typedi import Container
container = Container()

# Register both classes
container.register_class(A)
container.register_class(B)

a = container.resolve(A)
assert isinstance(a, A)
assert isinstance(a.b, B)

# Circular dependency
assert isinstance(a.b.a, A)
assert a.b.a == a 

assert a.b.a is not a  # NOTE: a.b.a is a proxy of a, not the exact instance

This code actually works. It is done internally by constructing a proxy object when encountering circular dependency during resolution. The proxy is completely identical to the original. You can use proxies in the same way as the original objects.

Proxy objects are used only to resolve a circular dependency. All non-circular dependencies would resolve in original, genuine objects.

The only restriction is that it is impossible to access the proxied circular dependencies inside the factory method (or __init__) itself. This code will raise an error during resolution:

...

class B:
    def __init__(self, a: A):            
        print(a)  # access to method __str__ of not-yet initialized proxy instance of A        
        self.a = a

...

Performance considerations

As you might have guessed already, there is a lot of introspection, reflection, type(...) and isinstance(...) going on under the hood when registering and resolving. I do not recommend using Container methods in your runtime hot paths since .resolve() call is quite expensive and might affect performance. Use it to assemble things on startup and throw the Container away.

Container tries its best to reduce unnecessary calls to factories. However, if return annotation of a provider is not strict enough (Union, or Any) in order to resolve some instance of that factory there is no other way other than call it and just go through all the options and check if it does match the query. Resolution complexity aligns in a way: ConcreteType < Union[ConcreteTypeA, ConcreteTypeB] < Any.

Another costly part is generators, if you have a generator that produces instance of types A, B and then C, resolving a C would also result in instantiating A and B:

from typing import Union, Iterable

class A: ...
class B: ...
class C: ...

def generator() -> Iterable[Union[A, B, C]]:
    yield A()  # wasted call
    yield B()  # wasted call
    yield C()
    yield A()  # also wasted call because container caches an entire result of a factory

from typedi import Container
container = Container()
container.register_factory(generator)

# generator would be called, and container will go through all generated instances 
# and check isinstance(item, C), once found, instance of C will be returned
c = container.resolve(C)
assert isinstance(c, C)

About python 3.6 and older

typing module became more or less stable only after python 3.7 (including). API for types in 3.6 differs significantly from 3.7 onward.

It is hard to maintain two different APIs that's why I decided to stick to the newer one. However, if you feel brave enough and really need that 3.6 support you can open a PullRequest and contact me if you have any questions.

Testing

This project is using tox (and pytest) to test among multiple python versions. To run test suites and generate coverage reports simply execute

tox

If you don't have tox installed, execute pip install tox first.