andi

Library for annotation-based dependency injection


License
BSD-3-Clause
Install
pip install andi==0.4.1

Documentation

andi

PyPI Version Supported Python Versions Build Status Coverage report

andi makes easy implementing custom dependency injection mechanisms where dependencies are expressed using type annotations.

andi is useful as a building block for frameworks, or as a library which helps to implement dependency injection (thus the name - ANnotation-based Dependency Injection).

License is BSD 3-clause.

Installation

pip install andi

andi requires Python >= 3.5.3.

Goal

See the following classes that represents parts of a car (and the car itself):

class Valves:
    pass

class Engine:
    def __init__(self, valves):
        self.valves = valves

class Wheels:
    pass

class Car:
    def __init__(self, engine, wheels):
        self.engine = engine
        self.wheels = wheels

The following would be the usual way of build a Car instance:

valves = Valves()
engine = Engine(valves)
wheels = Wheels()
car = Car(engine, wheels)

There are some dependencies between the classes: A car requires and engine and wheels to be built, as well as the engine requires valves. These are the car dependencies and sub-dependencies.

The question is, could we have an automatic way of building instances? For example, could we have a build function that given the Car class or any other class would return an instance even if the class itself has some other dependencies?

car = build(Car)  # Andi helps creating this generic build function

andi inspect the dependency tree and creates a plan making easy creating such a build function.

This is how this plan for the Car class would looks like:

  1. Invoke Valves with empty arguments
  2. Invoke Engine using the instance created in 1 as the argument valves
  3. Invoke Wheels with empty arguments
  4. Invoke Cars with the instance created in 2 as the engine argument and with the instance created in 3 as the wheels argument

Type annotations

But there is a missing piece in the Car example before. How can andi know that the class Valves is required to build the argument valves? A first idea would be to use the argument name as a hint for the class name (as pinject does), but andi opts to rely on arguments' type annotations instead.

The classes for Car should then be rewritten as:

class Valves:
    pass

class Engine:
    def __init__(self, valves: Valves):
        self.valves = valves

class Wheels:
    pass

class Car:
    def __init__(self, engine: Engine, wheels: Wheels):
        self.engine = engine
        self.wheels = wheels

Note how now there is a explicit annotation stating that the valves argument is of type Valves (same for engine and wheels).

The andi.plan function can now create a plan to build the Car class (ignore the is_injectable parameter by now):

plan = andi.plan(Car, is_injectable={Engine, Wheels, Valves})

This is what the plan variable contains:

[(Valves, {}),
 (Engine, {'valves': Valves}),
 (Wheels, {}),
 (Car,    {'engine': Engine,
           'wheels': Wheels})]

Note how this plan correspond exactly to the 4-steps plan described in the previous section.

Building from the plan

Creating a generic function to build the instances from a plan generated by andi is then very easy:

def build(plan):
    instances = {}
    for fn_or_cls, kwargs_spec in plan:
        instances[fn_or_cls] = fn_or_cls(**kwargs_spec.kwargs(instances))
    return instances

So let's see putting all the pieces together. The following code creates an instance of Car using andi:

plan = andi.plan(Car, is_injectable={Engine, Wheels, Valves})
instances = build(plan)
car = instances[Car]

is_injectable

It is not always desired for andi to manage every single annotation found. Instead is usually better to explicitly declare which types can be handled by andi. The argument is_injectable allows to customize this feature.

andi will raise an error on the presence of a dependency that cannot be resolved because it is not injectable.

Usually is desirable to declare injectabilty by creating a base class to inherit from. For example, we could create a base class Injectable as base class for the car components:

class Injectable(ABC):
    pass

class Valves(Injectable):
    pass

class Engine(Injectable):
    def __init__(self, valves: Valves):
        self.valves = valves

class Wheels(Injectable):
    pass

The call to andi.plan would then be:

is_injectable = lambda cls: issubclass(cls, Injectable)
plan = andi.plan(Car, is_injectable=is_injectable)

Functions and methods

Dependency injection is also very useful when applied to functions. Imagine that you have a function drive that drives the Car through the Road:

class Road(Injectable):
    ...

def drive(car: Car, road: Road, speed):
    ... # Drive the car through the road

The dependencies has to be resolved before invoking the drive function:

plan = andi.plan(drive, is_injectable=is_injectable)
instances = build(plan.dependencies)

Now the drive function can be invoked:

drive(instances[Car], instances[Road], 100)

Note that speed argument was not annotated. The resultant plan just won't include it because the andi.plan full_final_kwargs parameter is False by default. Otherwise, an exception would have been raised (see full_final_kwargs argument documentation for more information).

An alternative and more generic way to invoke the drive function would be:

drive(speed=100, **plan.final_kwargs(instances))

dataclasses and attrs

andi supports classes defined using attrs and also dataclasses. For example the Car class could have been defined as:

# attrs class example
@attr.s(auto_attribs=True)
class Car:
    engine: Engine
    wheels: Wheels

# dataclass example
@dataclass
class Car(Injectable):
    engine: Engine
    wheels: Wheels

Using attrs or dataclass is handy because they avoid some boilerplate.

Externally provided dependencies

Retaining the control over object instantiation could be desired in some cases. For example creating a database connection could require accessing some credentials registry or getting the connection from a pool so you might want to control building such instances outside of the regular dependency injection mechanism.

andi.plan allows to specify which types would be externally provided. Let's see an example:

class DBConnection(ABC):

    @abstractmethod
    def getConn():
        pass

@dataclass
class UsersDAO:
    conn: DBConnection

    def getUsers():
       return self.conn.query("SELECT * FROM USERS")

UsersDAO requires a database connection to run queries. But the connection will be provided externally from a pool, so we call then andi.plan using also the externally_provided parameter:

plan = andi.plan(UsersDAO, is_injectable=is_injectable,
                 externally_provided={DBConnection})

The build method should then be modified slightly to be able to inject externally provided instances:

def build(plan, instances_stock=None):
    instances_stock = instances_stock or {}
    instances = {}
    for fn_or_cls, kwargs_spec in plan:
        if fn_or_cls in instances_stock:
            instances[fn_or_cls] = instances_stock[fn_or_cls]
        else:
            instances[fn_or_cls] = fn_or_cls(**kwargs_spec.kwargs(instances))
    return instances

Now we are ready to create UserDAO instances with andi:

plan = andi.plan(UsersDAO, is_injectable=is_injectable,
                 externally_provided={DBConnection})
dbconnection = DBPool.get_connection()
instances = build(plan.dependencies, {DBConnection: dbconnection})
users_dao = instances[UsersDAO]
users = user_dao.getUsers()

Note that being injectable is not required for externally provided dependencies.

Optional

Optional type annotations can be used in case of dependencies that can be optional. For example:

@dataclass
class Dashboard:
    conn: Optional[DBConnection]

    def showPage():
        if self.conn:
            self.conn.query("INSERT INTO VISITS ...")
        ...  # renders a HTML page

In this example, the Dashboard class generates a HTML page to be served, and also stores the number of visits into a database. Database could be absent in some environments, but you might want the dashboard to work even if it cannot log the visits.

When a database connection is possible the plan call would be:

plan = andi.plan(UsersDAO, is_injectable=is_injectable,
                 externally_provided={DBConnection})

And the following when the connection is absent:

plan = andi.plan(UsersDAO, is_injectable=is_injectable,
                 externally_provided={})

It is also required to register the type of None as injectable. Otherwise andi.plan with raise an exception saying that "NoneType is not injectable".

Injectable.register(type(None))

Union

Union can also be used to express alternatives. For example:

@dataclass
class UsersDAO:
    conn: Union[ProductionDBConnection, DevelopmentDBConnection]

DevelopmentDBConnection will be injected in the absence of ProductionDBConnection.

Full final kwargs mode

By default andi.plan won't fail if it is not able to provide some of the direct dependencies for the given input (see the speed argument in one of the examples above).

This behaviour is desired when inspecting functions for which is already known that some arguments won't be injectable but they will be provided by other means (like the drive function above).

But in other cases is better to be sure that all dependencies are fulfilled and otherwise fail. Such is the case for classes. So it is recommended to set full_final_kwargs=True when invoking andi.plan for classes.

Why type annotations?

andi uses type annotations to declare dependencies (inputs). It has several advantages, and some limitations as well.

Advantages:

  1. Built-in language feature.
  2. You're not lying when specifying a type - these annotations still work as usual type annotations.
  3. In many projects you'd annotate arguments anyways, so andi support is "for free".

Limitations:

  1. Callable can't have two arguments of the same type.
  2. This feature could possibly conflict with regular type annotation usages.

If your callable has two arguments of the same type, consider making them different types. For example, a callable may receive url and html of a web page:

def parse(html: str, url: str):
    # ...

To make it play well with andi, you may define separate types for url and for html:

class HTML(str):
    pass

class URL(str):
    pass

def parse(html: HTML, url: URL):
    # ...

This is more boilerplate though.

Why doesn't andi handle creation of objects?

Currently andi just inspects callable and chooses best concrete types a framework needs to create and pass to a callable, without prescribing how to create them. This makes andi useful in various contexts - e.g.

  • creation of some objects may require asynchronous functions, and it may depend on libraries used (asyncio, twisted, etc.)
  • in streaming architectures (e.g. based on Kafka) inspection may happen on one machine, while creation of objects may happen on different nodes in a distributed system, and then actually running a callable may happen on yet another machine.

It is hard to design API with enough flexibility for all such use cases. That said, andi may provide more helpers in future, once patterns emerge, even if they're useful only in certain contexts.

Examples: callback based frameworks

Spider example

Nothing better than a example to understand how andi can be useful. Let's imagine you want to implemented a callback based framework for writing spiders to crawl web pages.

The basic idea is that there is framework in which the user can write spiders. Each spider is a collection of callbacks that can process data from a page, emit extracted data or request new pages. Then, there is an engine that takes care of downloading the web pages and invoking the user defined callbacks, chaining requests with its corresponding callback.

Let's see an example of an spider to download recipes from a cooking page:

class MySpider(Spider):
    start_url = "htttp://a_page_with_a_list_of_recipes"

    def parse(self, response):
        for url in recipes_urls_from_page(response)
            yield Request(url, callback=parse_recipe)

    def parse_recipe(self, response):
        yield extract_recipe(response)

It would be handy if the user can define some requirements just by annotating parameters in the callbacks. And andi make it possible.

For example, a particular callback could require access to the cookies:

def parse(self, response: Response, cookies: CookieJar):
    # ... Do something with the response and the cookies

In this case, the engine can use andi to inspect the parse method, and detect that Response and CookieJar are required. Then the framework will build them and will invoke the callback.

This functionality would serve to inject into the users callbacks some components only when they are required.

It could also serve to encapsulate better the user code. For example, we could just decouple the recipe extraction into it's own class:

@dataclass
class RecipeExtractor:
    response: Response

    def to_item():
        return extract_recipe(self.response)

The callback could then be defined as:

def parse_recipe(extractor: RecipeExtractor):
    yield extractor.to_item()

Note how handy is that with andi the engine can create an instance of RecipesExtractor feeding it with the declared Response dependency.

In definitive, using andi in such a framework can provide great flexibility to the user and reduce boilerplate.

Web server example

andi can be useful also for implementing a new web framework.

Let's imagine a framework where you can declare your sever in a class like the following:

class MyWeb(Server):

    @route("/products")
    def productspage(self, request: Request):
        ... # return the composed page

    @route("/sales")
    def salespage(self, request: Request):
        ... # return the composed page

The former case is composed of two endpoints, one for serving a page with a summary of sales, and a second one to serve the products list.

Connection to the database can be required to sever these pages. This logic could be encapsulated in some classes:

@dataclass
class Products:
    conn: DBConnection

    def get_products()
        return self.conn.query("SELECT ...")

@dataclass
class Sales:
    conn: DBConnection

    def get_sales()
        return self.conn.query("SELECT ...")

Now productspage and salespage methods can just declare that they require these objects:

class MyWeb(Server):

    @route("/products")
    def productspage(self, request: Request, products: Products):
        ... # return the composed page

    @route("/sales")
    def salespage(self, request: Request, sales: Sales):
        ... # return the composed page

And the framework can then be responsible to fulfill these dependencies. The flexibility offered would be a great advantage. As an example, if would be very easy to create a page that requires both sales and products:

@route("/overview")
def productspage(self, request: Request,
                 products: Products, sales: Sales):
    ... # return the composed overview page

Contributing

Use tox to run tests with different Python versions:

tox

The command above also runs type checks; we use mypy.