flexdi 0.1.4

A flexible dependency injection tool for typed Python projects.

FlexDI

flexdi is a yet another dependency injection library for Python. flexdi provides a lightweight alternative to common DI solutions with minimal setup to be included in your projects. This library is intended for use with type annotated Python libraries, as it leverages these type annotations to perform injection.

Note

FlexDI is still a work in progress and is not yet intended for production use-cases.

Be aware that APIs are likely to change in upcoming releases.

The full documentation is available at flexdi.readthedocs.io.

Goals

• Minimal Setup

  Some DI systems require that you conform to a particular style of class definition, a naming convention to arguments, placing wrappers around class or method definitions, or providing complex default arguments which are tied to the DI system. flexdi mainly uses type annotations to perform injection, which avoids a common problem of library lock-in where frameworks force you to design components using their highly specialized constructs.
• Inject Any Callable

  Basic Python DI tools only allow classes to be used for injection, which can be quite limiting. In flexdi, you can provide any typed callable as an input to be invoked. This avoids having to make class definitions purely for injection to work.
• Lifetime Management

  When creating dependencies for your callable, you often have resources that need to be properly torn down when your work is done. Providers for types in flexdi can be initialized using context managers that will have their shutdown logic invoked when you’re done your work.
• Asyncio Support

  flexdi allows calling both sync and async methods directly, and provides both a sync and async interface to its main invocation method. Similarly, dependency providers may be defined as async methods, or be provided as async context managers.

Overview

flexdi offers a construct called the FlexGraph which is used to keep track of dependencies and invoke other callables.

When determining dependencies for a callable, flexdi will examine the type annotations of the arguments, and populate the graph with dependencies which can satisfy the callable. A callable can be anything from a class (as seen with the type annotations), to functions, class methods, generators, etc.

For complex types, flexdi allows binding helper functions that can map a type definition to an instance. These bindings can themselves be injected with dependencies. Bindings can also be defined as generators which allows supplying custom teardown logic for dependencies.

Example Usage

A simple example of an application with SQLAlchemy dependencies:

import sys
from typing import Iterator
from sqlalchemy import Engine, create_engine, text
from sqlalchemy.orm import Session

from flexdi import FlexGraph

# The FlexGraph keeps track of what dependencies different
# providers require, and will later be used to resolve them.
graph = FlexGraph()

# Anything that requires an Engine will fetch it from provide_engine
# For simple functions we infer the binding from the return type annotation.
@graph.bind
def provide_engine() -> Engine:
    return create_engine("sqlite://")


# Generator responses can also be inferred. e.g.
# - A function returning Iterator[T] binds to T
# - A function returning AsyncIterator[T] binds to T
@graph.bind
def provide_session(engine: Engine) -> Iterator[Session]:
    with Session(engine) as session:
        yield session


# We don't need to add any flexdi setup to our actual code.
def main(session: Session) -> int:
    print(session.execute(text("SELECT now()")))
    return 0


if __name__ == "__main__":
    # Start up the injector, and guard using a with statement to
    # ensure that we clean up any dependencies which require it
    with graph:
        sys.exit(graph.resolve(main))

The same example, but using async code:

import sys
from typing import AsyncIterator
from sqlalchemy.ext.asyncio import (
    AsyncConnection,
    AsyncEngine,
    create_async_engine
)
from sqlalchemy import text

from flexdi import FlexGraph

graph = FlexGraph()


@graph.bind
async def provide_engine() -> AsyncIterator[AsyncEngine]:
    engine = create_async_engine("sqlite://")
    try:
        yield engine
    finally:
        await engine.dispose()


@graph.bind
async def provide_connection(engine: AsyncEngine) -> AsyncIterator[AsyncConnection]:
    async with engine.begin() as conn:
        yield conn


async def main(conn: AsyncConnection) -> int:
    print(await conn.execute(text("SELECT now()")))
    return 0


if __name__ == "__main__":
    with graph:
        # The injector can handle invoking async functions natively,
        # so no worry about adding in extra logic here.
        sys.exit(graph.resolve(main))
...


# If already within an async context, then you can use the
# async versions of these methods.
async def func() -> int:
    async with graph:
        return await graph.aresolve(main)

Alternatives

Although there are many, many other dependency injection libraries, I found that I was still left looking for more lightweight/minimal solutions to this problem. My thoughts on some of the popular alternatives I have used in the past:

• dependency-injector

  This library is probably the most mature out of all the alternatives. Its main driving principal is that “Explicit is better than implicit”, in that you need to specify explicitly how to assemble/ inject the dependencies. flexdi is still explicit in the sense that dependencies are directly referenced from their type annotations, and by leveraging them we can avoid a lot of the more verbose setup required in DeclarativeContainer structures.
• fastapi

  This web framework provides an excellent way to perform dependency injection, but it does not provide a way to perform dependency injection outside the context of web request. When configuring the injection, you must also provide default values to arguments, which ties application code to the web framework, making it more difficult to re-use code in other contexts. Additionally, it does not provide rich support for lifetime/singleton scoped dependencies, making the setup of some dependencies increasingly awkward.
• pinject

  This library allows you to perform DI with minimal setup, but its major downfall is that it relies on the names of arguments to perform injection. If the name of the argument does not match the name of the class, then you are forced to bind it explicitly. If there are multiple objects that specify a dependency of a particular type, but use different names, then you need to bind them all individually as well. And sadly, this project has now been archived and is read-only.