ididi 1.0.4

A dependency injection library for Python, Optimized for serverless applications

Ididi

Genius simplicity, unmathced power.

ididi is 100% test covered and strictly typed.

Information

Source Code ididi-github

Docs ididi-docs

Install

pip install ididi

To view viusal dependency graph, install graphviz

pip install ididi[graphviz]

Usage

Quick Start

Your existing code

import ididi

class UserService:
    def __init__(self, repo: UserRepository):
        self.repo = repo

user_service = ididi.resolve(UserService) 

Automatic dependencies injection

You can use generator/async generator to create a resource that needs to be closed. NOTE:

1. resources, if set to be reused, will be shared across different dependents only within the same scope, and destroyed when the scope is exited.
2. async resource in a sync dependent is not supported, but sync resource in a async dependent is supported.

from ididi import DependencyGraph

dg = DependencyGraph()

@dg.node
async def get_db(client: Client) -> ty.AsyncGenerator[DataBase, None]:
    db = DataBase(client)
    assert client.is_opened
    try:
        await db.connect()
        yield db
    finally:
        await db.close()

@dg.entry
async def main(db: DataBase, sql: str) -> ty.Any:
    res = await db.execute(sql)
    return res

assert await main(sql="select money from bank")

Using Scope to manage resources

• Infinite nested scope is supported.

• Parent scope can be accssed by child scope(within the same context)

• Resources will be shared across dependents only withint the same scope(reuse needs to be True)

• Resources will be automatically closed and destroyed when the scope is exited.

• Classes that implment contextlib.AbstractContextManager or contextlib.AbstractAsyncContextManager are also considered to be resources and can/should be resolved within scope.

• Scopes are separated by context

[!NOTE] If you have two call stack of a1 -> b1 and a2 -> b2, Here a1 and a2 are two calls to the same function a, then, in b1, you can only access scope created by the a1, not a2.

This is particularly useful when you try to separate resources by route, endpoint, request, etc.

Async, or not, works either way

@dg.node
def get_resource() -> ty.Generator[Resource, None, None]:
    res =  Resource()
    yield res
    res.close()

with dg.scope() as scope:
    resource = scope.resolve(Resource)

# For async generator
async with dg.scope() as scope:
    resource = await scope.resolve(Resource)

Contexted Scope

You can use dg.use_scope to retrive most recent scope, context-wise, this allows your to have access the scope without passing it around, e.g.

async def service_factory():
    async with dg.scope() as scope:
        service = scope.resolve(Service)
        yield service

@app.get("users")
async def get_user(service: Service = Depends(dg.factory(service_factory)))
    await service.create_user(...)

Then somewhere deep in your service.create_user call stack

async def create_and_publish():
    uow = dg.use_scope().resolve(UnitOfWork)
    async with uow.trans():
        user_repo.add_user(user)
        event_store.add(user_created_event)

Here dg.use_scope() would return the same scope you created in your service_factory.

Named Scope

You can create infinite level of scopes by assigning hashable name to scopes

# at the top most entry of a request
async with dg.scope(request_id) as scope:
    ...

now scope with name request_id is accessible everywhere within the request context

request_scope = dg.use_scope(request_id)

[!NOTE] Two scopes or more with the same name would follow most recent rule.

Nested Nmaed Scope

async with dg.scope(app_name) as app_scope:
    async with dg.scope(router) as router_scope:
        async with dg.scope(endpoint) as endpoint_scope:
            async with dg.scope(user_id) as user_scope:
                async with dg.scope(request_id) as request_scope:
                    ...

For any functions called within the request_scope, you can get the most recent scope with dg.use_scope(), or its parent scopes, i.e. dg.use_scope(app_name) to get app_scope.

Usage with FastAPI

from fastapi import FastAPI
from ididi import DependencyGraph

app = FastAPI()
dg = DependencyGraph()

def auth_service_factory(db: DataBase) -> AuthService:
    async with dg.scope() as scope
        yield dg.resolve(AuthService)

Service = ty.Annotated[AuthService, Depends(dg.factory(auth_service_factory))]

@app.get("/")
def get_service(service: Service):
    return service

Visualize the dependency graph(beta)

from ididi import DependencyGraph, Visualizer
dg = DependencyGraph()

class ConfigService:
    def __init__(self, env: str = "test"):
        self.env = env


class DatabaseService:
    def __init__(self, config: ConfigService):
        self.config = config


class CacheService:
    def __init__(self, config: ConfigService):
        self.config = config


class BaseService:
    def __init__(self, db: DatabaseService):
        self.db = db


class AuthService(BaseService):
    def __init__(self, db: DatabaseService, cache: CacheService):
        super().__init__(db)
        self.cache = cache


class UserService:
    def __init__(self, auth: AuthService, db: DatabaseService):
        self.auth = auth
        self.db = db


class NotificationService:
    def __init__(self, config: ConfigService):
        self.config = config


class EmailService:
    def __init__(self, notification: NotificationService, user: UserService):
        self.notification = notification
        self.user = user

dg.static_resolve(EmailService)
vs = Visualizer(dg)
vs.view # use vs.view in jupyter notebook, or use vs.save(path, format) otherwise
vs.save(path, format)

Circular Dependency Detection

ididi would detect if circular dependency exists, if so, ididi would give you the circular path

For example:

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


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


class C:
    def __init__(self, d: "D"):
        pass


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


def test_advanced_cycle_detection():
    """
    DependentNode.resolve_forward_dependency
    """
    dag = DependencyGraph()

    with pytest.raises(CircularDependencyDetectedError) as exc_info:
        dag.static_resolve(A)
    assert exc_info.value.cycle_path == [A, B, C, D, A]

You can call DependencyGraph.static_resolve_all on app start to statically resolve all your noded classes, and let ididi get ready for resolve them at upcoming calls.

Lazy Dependency(Beta)

you can use @dg.node(lazy=True) to define a dependent as lazy, which means each of its dependency will not be resolved untill accessed.

start with v0.3.0, lazy node is no longer transitive.

class UserRepo:
    def __init__(self, db: Database):
        self._db = db

    def test(self):
        return "test"

@dg.node(lazy=True)
class ServiceA:
    def __init__(self, user_repo: UserRepo, session_repo: SessionRepo):
        self._user_repo = user_repo
        self._session_repo = session_repo

        assert isinstance(self._user_repo, LazyDependent)
        assert isinstance(self._session_repo, LazyDependent)

    @property
    def user_repo(self) -> UserRepo:
        return self._user_repo

    @property
    def session_repo(self) -> SessionRepo:
        return self._session_repo

assert isinstance(instance.user_repo, LazyDependent)
assert isinstance(instance.session_repo, LazyDependent)

# user_repo would be resolved when user_repo.test() is called.
assert instance.user_repo.test() == "test" 

Runtime override

dg = DependencyGraph()

class Inner:
    def __init__(self, value: str = "inner"):
        self.value = value

@dg.node
class Outer:
    def __init__(self, inner: Inner):
        self.inner = inner

# Override nested dependency
instance = dg.resolve(Outer, inner=Inner(value="overridden"))
assert instance.inner.value == "overridden"

Advanced Usage

ABC

Register ABC implementation with dg.node

you should use dg.node to let ididi know about the implementations of the ABC. you are going to resolve.

from abc import ABC, abstractmethod
class Repository(ABC):
    def __init__(self):
        pass

    @abstractmethod
    def save(self) -> None:
        """Save the repository data."""
        pass

@dag.node
class Repo1(Repository):
    def save(self) -> None:
        pass

@dag.node
class Repo2(Repository):
    def save(self) -> None:
        pass

dag.resolve(Repository)

You might also use __init_subclass__ hook to automatically register implementations.

Multiple Implementations of ABC

ididi will use the last implementation registered to resolve the ABC, you can use a factory to override this behavior.

class Repository(ABC):
    def __init__(self):
        pass

    @abstractmethod
    def save(self) -> None:
        """Save the repository data."""
        pass

@dag.node
class Repo1(Repository):
    def save(self) -> None:
        pass

@dag.node
class Repo2(Repository):
    def save(self) -> None:
        pass

@dag.node
def repo_factory() -> Repository:
    return Repo1()

assert Repository in dag.nodes

repo = dag.resolve(Repository)
assert isinstance(repo, Repo1)

Dependent that implements multiple protocols

class UserRepo(ty.Protocol): ...

class SessionRepo(ty.Protocol): ...

@dg.node
class BothRepo(UserRepo, SessionRepo):
    def __init__(self, name: str = "test"):
        self.name = name

class UserApp:
    def __init__(self, repo: UserRepo):
        self.repo = repo

class SessionApp:
    def __init__(self, repo: SessionRepo):
        self.repo = repo

user = scope.resolve(UserApp)
session = scope.resolve(SessionApp)
assert user.repo is session.repo

# same logic for resource with scope

Performance

ididi is very efficient and performant, with average time complexity of O(n)

DependencyGraph.statical_resolve (type analysis on each class, can be done at import time)

Time Complexity: O(n) - O(n**2)

O(n): average case, where each dependent has a constant number of dependencies, for example, each dependents has on average 3 dependencies.

O(n**2): worst case, where each dependent has as much as possible number of dependencies, for example, with 100 nodes, node 1 has 99 dependencies, node 2 has 98, etc.

I personally don't think anyone would ever encouter the worse case in real-world, but even if someone does, you can still expect ididi resolve thousand of such classes in seoncds.

As a reference:

tests/test_benchmark.py 0.003801 seoncds to statically resolve 122 classes

DependencyGraph.resolve (inject dependencies and build the dependent instance)

Time Complexity: O(n)

You might run the benchmark yourself with following steps

1. clone the repo, cd to project root
2. install pixi from pixi
3. run pixi install
4. run make benchmark

Performance tip

• use dg.node to decorate your classes

• use dg.node to decorate factory of third party classes so that ididi does not need to analyze them

For Example

def redis_factory(settings: Settings) -> Redis:
    # build redis here
    return redis

• use dg.static_resolve_all when your app starts, which will statically resolve all your classes decorated with @dg.node.

Resolve Rules

• If a node has a factory, it will be used to create the instance.
• Otherwise, the node will be created using the __init__ method.
  • Parent's __init__ will be called if no __init__ is defined in the node.
• whenver there is a default value, it will be used to resolve the dependency.
• bulitin types are not resolvable by nature, it requires default value to be provided.
• runtime override with dg.resolve

Error context

static resolve might fail when class contain unresolvable dependencies, when failed, ididi would show a chain of errors like this:

ididi.errors.MissingAnnotationError: Unable to resolve dependency for parameter: env in <class 'tests.features.test_improved_error.Config'>, annotation for `env` must be provided

<- Config(env: _Missed)
<- DataBase(config: Config)
<- AuthService(db: DataBase)
<- UserService(auth: AuthService)

Where UserService depends on AuthService, which depends on Database, then Config, but Config.env is missing annotation.

What and why

What is dependency injection?

If a class requires other classes as its attributes, then these attributes are regarded as dependencies of the class, and the class requiring them is called a dependent.

class Downloader:
    def __init__(self, session: requests.Session):
        self.session = session

Here, Downloader is a dependent, with requests.Session being its dependency.

Dependency injection means dynamically constructing the instances of these dependency classes and then pass them to the dependent class.

the same class without dependency injection looks like this:

class Downloader:
    def __init__(self):
        self.session = requests.Session(url=configured_url, timeout=configured_timeout)

Now, since requests.Session is automatically built with Downloader, it would be difficult to change the behavior of requests.Session at runtime.

Why do we need it?

There are actually a few reasons why you might not need it, the most fundamental one being your code does not need reuseability and flexibility.

1. If you are writing a script that only runs when you menually execute it, and it is often easier to rewrite the whole script than to modify it, then it probably more efficient to program everything hard-coded. This is actually a common use case of python, DEVOPS, DataAnalysts, etc.

For example, you can actually modify the dependencies of a class at runtime.

class Downloader:
    ...

downloader = Downloader()
downloader.session = requests.Session(url=configured_url, timeout=configured_timeout)

However, this creates a few problems:

• It is error-prone, you might forget to modify the dependencies, or you might modify the dependencies in the wrong order.
• It is not typesafe, you might pass the wrong type of dependencies to the class.
• It is hard to track when the dependencies are modified.

Dependency injection enables you to extend the dependencies of a class without modifying the class itself, which increases the flexibility and reusability of the class.

Do we need a DI framework?

Not necessarily, You will be doing just fine using menual dependency injection,as long as the number of dependencies in your app stays within a managable range.

It gets more and more helpful once your your dependency graph starts getting more complicated, For example, you might have something like this in your app,

where you menually inject dependencies into dependent.

from .infra import * 

def auth_service_factory(
    settings: Settings,
) -> AuthService:
    connect_args = (
        settings.db.connect_args.model_dump()
    )
    execution_options = (
        settings.db.execution_options.model_dump()
    )
    engine = engine_factory(
        db_url=settings.db.DB_URL,
        echo=settings.db.ENGINE_ECHO,
        isolation_level=settings.db.ISOLATION_LEVEL,
        pool_pre_ping=True,
        connect_args=connect_args,
        execution_options=execution_options,
    )
    async_engine = sqlalchemy.ext.asyncio.AsyncEngine(engine)
    db = AsyncDatabase(async_engine)
    cache = RedisCache[str].build(
        url=config.URL,
        keyspace=config.keyspaces.APP,
        socket_timeout=config.SOCKET_TIMEOUT,
        decode_responses=config.DECODE_RESPONSES,
        max_connections=config.MAX_CONNECTIONS,
    )
    token_registry = TokenRegistry(cahce=cache, 
        token_bucket=TokenBucket(cache, key=Settings.redis.bucket_key)
    )
    uow = UnitOfWork(db)
    encryptor = Encryptor(
        secret_key=settings.security.SECRET_KEY.get_secret_value(),
        algorithm=settings.security.ALGORITHM,  
    )
    eventstore =  EventStore(uow)
    auth_repo = AuthRepository(db)
    auth_service = AuthService(
        auth_repo=auth_repo,
        token_registry=token_registry,
        encryptor=encryptor,
        eventstore=eventstore,
        security_settings=settings.security,
    )
    return auth_service

But then you realize that some of these dependencies should be shared across your services, for example, auth repo might be needed by both AuthService and UserService, or even more

You might also need to menually create and manage scope as some resources should be accessed/shared only whtin a certain scope, e.g., a request.

Why Ididi?

ididi helps you do this while stays out of your way, you do not need to create additional classes like Container, Provider, Wire, nor adding lib-specific annotation like Closing, Injectable, etc.

ididi provides unique powerful features that most alternatives don't have, such as support to inifinite number of context-specific nested sopce, lazydependent, advanced circular dependency detection, plotting, etc.

Terminology

dependent: a class, or a function that requires arguments to be built/called.

dependency: an object that is required by a dependent.

resource: a dependent that implements the contextlib.AbstractAsync/ContextManager, or has an async/sync generator as its factory, is considered a resource.

static resolve: resursively build node from dependent, but does not create the instance of the dependent type.

resolve: recursively resolve the dependent and its dependencies, then create an instance of the dependent type.

solve: an alias for resolve

entry: a special type of node, where it has no dependents and its factory is itself.

FAQ

How do I override, or provide a default value for a dependency?

you can use dg.node to create a factory to override the value. you can also have dependencies in your factory, and they will be resolved recursively.

class Config:
    def __init__(self, env: str = "prod"):
        self.env = env

@dg.node
def config_factory() -> Config:
    return Config(env="test")

How do i override a dependent in test?

you can use dg.node with a factory method to override the dependent resolution.

class Cache: ...

class RedisCache(Cache):
    ...

class MemoryCache(Cache):
    ...

@dg.node
def cache_factory(...) -> Cache:
    return RedisCache() 

in your conftest.py:

@dg.node
def memory_cache_factory(...) -> Cache:
    return MemoryCache()    

as this follows LSP, it works both with ididi and type checker.

How do I make ididi reuse a dependencies across different dependent?

by default, ididi will reuse the dependencies across different dependent, you can change this behavior by setting reuse=False in dg.node.

@dg.node(reuse=False) # True by default
class AuthService: ...