aspyx 1.9.0

A DI and AOP library for Python

aspyx

Table of Contents

• Motivation
• Overview
• Installation
• Registration
  • Class
  • Class Factory
  • Method
  • Conditional
• Environment
  • Definition
  • Retrieval
• Instantiation logic
  • Injection Methods
  • Lifecycle Methods
  • Post Processors
• Custom scopes
• AOP
• Threading
• Configuration
• Reflection
• Exceptions
• Version History

Motivation

While working on AI-related projects in Python, I was looking for a dependency injection (DI) framework. After evaluating existing options, my impression was that the most either lacked key features — such as integrated AOP — or had APIs that felt overly technical and complex, which made me develop a library on my own with the following goals

• bring both di and AOP features together in a lightweight library,
• be as minimal invasive as possible,
• offering mechanisms to easily extend and customize features without touching the core,
• while still offering a simple and readable api that doesnt overwhelm developers and only requires a minimum initial learning curve

The AOP integration, in particular, makes a lot of sense because:

• Aspects typically require context, which is naturally provided through DI,
• And they should only apply to objects managed by the container, rather than acting globally.

Overview

Aspyx is a lightweight - still only about 2K LOC - Python library that provides both Dependency Injection (DI) and Aspect-Oriented Programming (AOP) support.

The following DI features are supported

• constructor and setter injection
• injection of configuration variables
• possibility to define custom injections
• post processors
• support for factory classes and methods
• support for eager and lazy construction
• support for scopes "singleton", "request" and "thread"
• possibility to add custom scopes
• conditional registration of classes and factories ( aka profiles in spring )
• lifecycle events methods on_init, on_destroy, on_running
• Automatic discovery and bundling of injectable objects based on their module location, including support for recursive imports
• Instantiation of one or possible more isolated container instances — called environments — each managing the lifecycle of a related set of objects,
• Support for hierarchical environments, enabling structured scoping and layered object management.

With respect to AOP:

• support for before, around, after and error aspects
• simple fluent interface to specify which methods are targeted by an aspect
• sync and async method support

The library is thread-safe and heavily performance optimized as most of the runtime information is precomputed and cached!

Let's look at a simple example

from aspyx.di import injectable, on_init, on_destroy, module, Environment


@injectable()
class Foo:
    def hello(self, msg: str):
        print(f"hello {msg}")


@injectable()  # eager and singleton by default
class Bar:
    def __init__(self, foo: Foo):  # will inject the Foo dependency
        self.foo = foo

    @on_init()  # a lifecycle callback called after the constructor and all possible injections
    def init(self):
        ...


# this class will discover and manage all - specifically decorated - classes and factories that are part of the own module

@module()
class SampleModule:
    pass

# create environment

environment = Environment(SampleModule)

# fetch an instance

bar = environment.get(Bar)

bar.foo.hello("world")

The concepts should be pretty familiar as well as the names as they are inspired by both Spring and Angular.

Let's add some aspects...

@advice
class SampleAdvice:
    def __init__(self): # could inject additional stuff
        pass

    @before(methods().named("hello").of_type(Foo))
    def call_before(self, invocation: Invocation):
        ...

    @error(methods().named("hello").of_type(Foo))
    def call_error(self, invocation: Invocation):
        ... # exception accessible in invocation.exception

    @around(methods().named("hello"))
    def call_around(self, invocation: Invocation):
        ...
        return invocation.proceed()

While features like DI and AOP are often associated with enterprise applcations, this example hopefully demonstrates that they work just as well in small- to medium-sized projects—without introducing significant overhead—while still providing powerful tools for achieving clean architecture, resulting in maintainable and easily testable code.

Let's look at the details

Installation

Just install from PyPI with

pip install aspyx

The library is tested with all Python version >= 3.9

Registration

Different mechanisms are available that make classes eligible for injection

Class

Any class annotated with @injectable is eligible for injection

Example:

@injectable()
class Foo:
    def __init__(self):
        pass

If the class defines a constructor, all parameters - which are expected to be registered as well - will be injected automatically.

The decorator accepts the keyword arguments

• eager : boolean
  if True, the container will create the instances automatically while booting the environment. This is the default.
• scope: str
  the name of a - registered - scope which will determine how often instances will be created.

The following scopes are implemented out of the box:

• singleton
  objects are created once inside an environment and cached. This is the default.
• request
  objects are created on every injection request
• thread
  objects are created and cached with respect to the current thread.

Other scopes - e.g. session related scopes - can be defined dynamically. Please check the corresponding chapter.

Class Factory

Classes that implement the Factory base class and are annotated with @factory will register the appropriate classes returned by the create method.

Example:

@factory()
class TestFactory(Factory[Foo]):
    def create(self) -> Foo:
        return Foo()

As in @injectable, the same arguments are possible.

Method

Any injectable can define methods decorated with @create(), that will create appropriate instances.

Example:

@injectable()
class Foo:
    @create(scope="request")
    def create(self) -> Baz:
        return Baz()

The same arguments as in @injectable are possible.

Conditional

All @injectable declarations can be supplemented with

@conditional(<condition>, ..., <condition>)

decorators that act as filters in the context of an environment.

Valid conditions are created by:

• requires_class(clazz: Type)
  the injectable is valid, if the specified class is registered as well.
• requires_feature(feature: str)
  the injectable is valid, if the environment defines the specified feature.

Environment

Definition

An Environment is the container that manages the lifecycle of objects. The set of classes and instances is determined by a constructor type argument called module.

Example:

@module()
class SampleModule:
    pass

A module is a regular injectable class decorated with @module that controls the discovery of injectable classes, by filtering classes according to their module location relative to this class. All eligible classes, that are implemented in the containing module or in any submodule will be managed.

In a second step the real container - the environment - is created based on a module:

environment = Environment(SampleModule, features=["dev"])

By adding the parameter features: list[str], it is possible to filter injectables by evaluating the corresponding @conditional decorators.

Example:

@injectable()
@conditional(requires_feature("dev"))
class DevOnly:
     pass

@module()
class SampleModule():
    pass

environment = Environment(SampleModule, features=["dev"])

By adding an imports: list[Type] parameter, specifying other module types, it will register the appropriate classes recursively.

Example:

@module()
class SampleModule(imports=[OtherModule]):
    pass

Another possibility is to add a parent environment as an Environment constructor parameter

Example:

rootEnvironment = Environment(RootModule)

environment = Environment(SampleModule, parent=rootEnvironment)

The difference is, that in the first case, class instances of imported modules will be created in the scope of the own environment, while in the second case, it will return instances managed by the parent.

The method

shutdown()

is used when a container is not needed anymore. It will call any on_destroy() of all created instances.

Retrieval

def get(type: Type[T]) -> T

is used to retrieve object instances. Depending on the respective scope it will return either cached instances or newly instantiated objects.

The container knows about class hierarchies and is able to get base classes, as long as there is only one implementation.

In case of ambiguities, it will throw an exception.

Note that a base class are not required to be annotated with @injectable, as this would mean, that it could be created on its own as well. ( Which is possible as well, btw. )

Instantiation logic

Constructing a new instance involves a number of steps executed in this order

• Constructor call
  the constructor is called with the resolved parameters
• Advice injection
  All methods involving aspects are updated
• Lifecycle methods
  different decorators can mark methods that should be called during the lifecycle ( here the construction ) of an instance. These are various injection possibilities as well as an optional final on_init call
• PostProcessors
  Any custom post processors, that can add side effects or modify the instances

Injection methods

Different decorators are implemented, that call methods with computed values

• @inject
  the method is called with all resolved parameter types ( same as the constructor call)
• @inject_environment
  the method is called with the creating environment as a single parameter
• @inject_value()
  the method is called with a resolved configuration value. Check the corresponding chapter

Example:

@injectable()
class Foo:
    @inject_environment()
    def set_environment(self, env: Environment):
        ...

    @inject()
    def set(self, baz: Baz) -> None:
        ...

Lifecycle methods

It is possible to mark specific lifecyle methods.

• @on_init() called after the constructor and all other injections.
• @on_running() called after an environment has initialized completely ( e.g. created all eager objects ).
• @on_destroy() called during shutdown of the environment

Post Processors

As part of the instantiation logic it is possible to define post processors that execute any side effect on newly created instances.

Example:

@injectable()
class SamplePostProcessor(PostProcessor):
    def process(self, instance: object, environment: Environment):
        print(f"created a {instance}")

Any implementing class of PostProcessor that is eligible for injection will be called by passing the new instance.

Note that a post processor will only handle instances after its own registration.

As injectables within a single file will be handled in the order as they are declared, a post processor will only take effect for all classes after its declaration!

Custom scopes

As explained, available scopes are "singleton" and "request".

It is easily possible to add custom scopes by inheriting the base-class Scope, decorating the class with @scope(<name>) and overriding the method get

def get(self, provider: AbstractInstanceProvider, environment: Environment, argProvider: Callable[[],list]):

Arguments are:

• provider the actual provider that will create an instance
• environmentthe requesting environment
• argProvider a function that can be called to compute the required arguments recursively

Example: The simplified code of the singleton provider ( disregarding locking logic )

@scope("singleton")
class SingletonScope(Scope):
    # constructor

    def __init__(self):
        super().__init__()

        self.value = None

    # override

    def get(self, provider: AbstractInstanceProvider, environment: Environment, argProvider: Callable[[],list]):
        if self.value is None:
            self.value = provider.create(environment, *argProvider())

        return self.value

AOP

It is possible to define different aspects, that will be part of method calling flow. This logic fits nicely in the library, since the DI framework controls the instantiation logic and can handle aspects within a regular post processor.

On the other hand, advices are also regular DI objects, as they will usually require some kind of - injected - context.

Advices are regular classes decorated with @advice that define aspect methods.

@advice
class SampleAdvice:
    def __init__(self):  # could inject dependencies
        pass

    @before(methods().named("hello").of_type(Foo))
    def call_before(self, invocation: Invocation):
        # arguments: invocation.args and invocation.kwargs
        ...

     @after(methods().named("hello").of_type(Foo))
    def call_after(self, invocation: Invocation):
        # arguments: invocation.args and invocation.kwargs
        ...

    @error(methods().named("hello").of_type(Foo))
    def call_error(self, invocation: Invocation):
         # error: invocation.exception
        ...

    @around(methods().named("hello"))
    def call_around(self, invocation: Invocation):
        try:
            ...
            return invocation.proceed()  # will leave a result in invocation.result or invocation.exception in case of an exception
        finally:
            ...

Different aspects - with the appropriate decorator - are possible:

• before
  methods that will be executed prior to the original method
• around
  methods that will be executed around to the original method allowing you to add side effects or even modify parameters.
• after
  methods that will be executed after to the original method
• error
  methods that will be executed in case of a caught exception

The different aspects can be supplemented with an @order(<prio>) decorator that controls the execution order based on the passed number. Smaller values get executed first.

All methods are expected to have single Invocation parameter, that stores

• func the target function
• args the supplied args ( including the self instance as the first element)
• kwargs the keywords args
• result the result ( initially None)
• exception a possible caught exception ( initially None)

⚠️ Note: It is essential for around methods to call proceed() on the invocation, which will call the next around method in the chain and finally the original method.

If the proceed is called with parameters, they will replace the original parameters!

Example: Parameter modifications

@around(methods().named("say"))
def call_around(self, invocation: Invocation):
    return invocation.proceed(invocation.args[0], invocation.args[1] + "!") # 0 is self!

The argument list to the corresponding decorators control which methods are targeted by the advice.

A fluent interface is used describe the mapping. The parameters restrict either methods or classes and are constructed by a call to either methods() or classes().

Both add the fluent methods:

• of_type(type: Type)
  defines the matching classes
• named(name: str)
  defines method or class names
• that_are_async()
  defines async methods
• matches(re: str)
  defines regular expressions for methods or classes
• decorated_with(type: Type)
  defines decorators on methods or classes

The fluent methods named, matches and of_type can be called multiple times!

Example: react on both transactional decorators on methods or classes

@advice
class TransactionAdvice:
    def __init__(self):
        pass

    @around(methods().decorated_with(transactional), classes().decorated_with(transactional))
    def establish_transaction(self, invocation: Invocation):
        ...

With respect to async methods, you need to make sure, to replace a proceed() with a await proceed_async() to have the overall chain async!

Advice Lifecycle and visibility.

Advices are always part of a specific environment, and only modify methods of objects managed by exactly this environment.

An advice of a parent environment will for example not see classes of inherited environments. What is done instead, is to recreate the advice - more technically speaking, a processor that will collect and apply the advices - in every child environment, and let it operate on the local objects. With this approach different environments are completely isolated from each other with no side effects whatsoever.

Threading

A handy decorator @synchronized in combination with the respective advice is implemented that automatically synchronizes methods with a RLock associated with the instance.

Example:

@injectable()
class Foo:
    @synchronized()
    def execute_synchronized(self):
        ...

Configuration

It is possible to inject configuration values, by decorating methods with @inject-value(<name>) given a configuration key.

@injectable()
class Foo:
    @inject_value("HOME")
    def inject_home(self, os: str):
        ...

If required type coercion will be applied.

Configuration values are managed centrally using a ConfigurationManager, which aggregates values from various configuration sources that are defined as follows.

class ConfigurationSource(ABC):
    @inject()
    def set_manager(self, manager: ConfigurationManager):
        manager._register(self)

    @abstractmethod
    def load(self) -> dict:
       pass

The load method is able to return a tree-like structure by returning a dict.

Configuration variables are retrieved with the method

def get(self, path: str, type: Type[T], default : Optional[T]=None) -> T:

• path
  a '.' separated path
• type
  the desired type
• default
  a default, if no value is registered

Sources can be added dynamically by registering them.

Example:

@injectable()
class SampleConfigurationSource(ConfigurationSource):
    def __init__(self):
        super().__init__()

    def load(self) -> dict:
        return {
            "a": 1, 
            "b": {
                "d": "2", 
                "e": 3, 
                "f": 4
                }
            }

Two specific source are already implemented:

• EnvConfigurationSource
  reads the os environment variables
• YamlConfigurationSource
  reads a specific yaml file

Typically you create the required configuration sources in an environment class, e.g.

@module()
class SampleModule:
    @create()
    def create_env_source(self) -> EnvConfigurationSource:
        return EnvConfigurationSource()

    @create()
    def create_yaml_source(self) -> YamlConfigurationSource:
        return YamlConfigurationSource("config.yaml")

Reflection

As the library heavily relies on type introspection of classes and methods, a utility class TypeDescriptor is available that covers type information on classes.

After being instantiated with

TypeDescriptor.for_type(<type>)

it offers the methods

• get_methods(local=False)
  return a list of either local or overall methods
• get_method(name: str, local=False)
  return a single either local or overall method
• has_decorator(decorator: Callable) -> bool
  return True, if the class is decorated with the specified decorator
• get_decorator(decorator) -> Optional[DecoratorDescriptor]
  return a descriptor covering the decorator. In addition to the callable, it also stores the supplied args in the args property

The returned method descriptors provide:

• param_types
  list of arg types
• return_type
  the return type
• has_decorator(decorator: Callable) -> bool return True, if the method is decorated with the specified decorator
• get_decorator(decorator: Callable) -> Optional[DecoratorDescriptor]
  return a descriptor covering the decorator. In addition to the callable, it also stores the supplied args in the args property

The management of decorators in turn relies on another utility class Decorators that caches decorators.

Whenver you define a custom decorator, you will need to register it accordingly.

Example:

def transactional(scope):
    def decorator(func):
        Decorators.add(func, transactional, scope) # also add _all_ parameters in order to cache them
        return func

    return decorator

Exceptions

The class ExceptionManager is used to collect dynamic handlers for specific exceptions and is able to dispatch to the concrete functions given a specific exception.

The handlers are declared by annoting a class with @exception_handler and decorating specific methods with @catch

Example:

class DerivedException(Exception):
    def __init__(self):
        pass


@module()
class SampleModule:
    @create()
    def create_exception_manager(self) -> ExceptionManager:
        return ExceptionManager()


@injectable()
@exception_handler()
class TestExceptionHandler:
    @catch()
    def catch_derived_exception(self, exception: DerivedException):
        ExceptionManager.proceed()

    @catch()
    def catch_exception(self, exception: Exception):
        pass

    @catch()
    def catch_base_exception(self, exception: BaseException):
        pass


@advice
class ExceptionAdvice:
    def __init__(self, exceptionManager: ExceptionManager):
        self.exceptionManager = exceptionManager

    @error(methods().of_type(Service))
    def handle_error(self, invocation: Invocation):
        self.exceptionManager.handle(invocation.exception)


environment = Environment(SampleEnvironment)

environment.read(ExceptionManager).handle(DerivedException())

The exception maanger will first call the most appropriate method. Any ExceptionManager.proceed() will in turn call the next most applicable method ( if available).

Together with a simple around advice we can now add exception handling to any method:

Example:

@injectable()
class Service:
    def throw(self):
        raise DerivedException()

@advice
class ExceptionAdvice:
    def __init__(self, exceptionManager: ExceptionManager):
        self.exceptionManager = exceptionManager

    @error(methods().of_type(Service))
    def handle_error(self, invocation: Invocation):
        self.exceptionManager.handle(invocation.exception)

Version History

1.0.1

• some internal refactorings

1.1.0

• added @on_running() callback
• added thread scope

1.2.0

• added YamlConfigurationSource

1.3.0

• added @conditional
• added support for async advices

1.4.0

• bugfixes
• added @ExceptionManager

1.4.1

• mkdocs

1.6.1

• default constructors not requires anymore