Skip to content

suned/stateless

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

48 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

stateless

Statically typed, purely functional effects for Python.

Motivation

Programming with side-effects is hard: To reason about a unit in your code, like a function, you need to know what the other units in the program are doing to the program state, and understand how that affects what you're trying to achieve.

Programming without side-effects is less hard: To reason about a unit in you code, like a function, you can focus on what that function is doing, since the units it interacts with don't affect the state of the program in any way.

But of course side-effects can't be avoided, since what we ultimately care about in programming are just that: The side effects, such as printing to the console or writing to a database.

Functional effect systems like stateless aim to make programming with side-effects less hard. We do this by separating the specification of side-effects from the interpretation, such that functions that need to perform side effects do so indirectly via the effect system.

As a result, "business logic" code never performs side-effects, which makes it easier to reason about, test and re-use.

Quickstart

from typing import Any, Never
from stateless import Effect, depend, throws, catch, Runtime


# stateless.Effect is just an alias for:
#
# from typing import Generator, Any
#
# type Effect[A, E: Exception, R] = Generator[Type[A] | E, Any, R]


class Files:
    def read_file(self, path: str) -> str:
        with open(path) as f:
            return f.read()


class Console:
    def print(self, value: Any) -> None:
        print(value)


# Effects are generators that yield "Abilities" that can be sent to the
# generator when an effect is executed. Abilities could be anything, but will often be things that
# handle side-effects. Here it's a class that can print to the console.
# In other effects systems, abilities are called "effect handlers".
def print_(value: Any) -> Effect[Console, Never, None]:
    console = yield from depend(Console)  # depend returns abilities
    console.print(value)


# Effects can yield exceptions. 'stateless.throws' will catch exceptions
# for you and yield them to other functions so you can handle them with
# type safety. The return type of the decorated function in this
# example is: Β΄Effect[Files, OSError, str]'
@throws(OSError)
def read_file(path: str) -> Effect[Files, Never, str]:
    files = yield from depend(Files)
    return files.read_file(path)


# Simple effects can be combined into complex ones by
# depending on multiple abilities.
def print_file(path: str) -> Effect[Files | Console, Never, None]:
    # catch will return exceptions yielded by other functions
    result = yield from catch(read_file)(path)
    match result:
        case OSError() as error:
            yield from print_(f"error: {error}")
        case _ as content:
            yield from print_(content)


# Effects are run using `stateless.Runtime.run`.
# Abilities are provided to effects via 'stateless.Runtime.use'
runtime = Runtime().use(Console()).use(Files())
runtime.run(print_file('foo.txt'))

Guide

Effects, Abilities and Runtime

stateless is a functional effect system for Python built around a pattern using generator functions. When programming with stateless you will describe your program's side-effects using the stateless.Effect type. This is in fact just a type alias for a generator:

from typing import Any, Generator, Type


type Effect[A, E: Exception, R] = Generator[Type[A] | E, Any, R]

In other words, an Effect is a generator that can yield classes of type A or exceptions of type E, can be sent anything, and returns results of type R. Let's break that down a bit further:

  • The type variable A in Effect stands for "Ability". This is the type of value that an effect depends on in order to produce its result.

  • The type variable E parameter of Effect stands for "Error". This the type of errors that an effect might fail with.

  • The type variable R stands for "Result". This is the type of value that an Effect will produce if no errors occur.

We'll see shortly why the "send" type of effects must be Any, and how stateless can still provide good type inference.

Lets start with a very simple example of an Effect:

from typing import Never

from stateless import Effect


def hello_world() -> Effect[str, Never, None]:
    message = yield str
    print(message)

When hello_world returns an Effect[str, Never, None], it means that it depends on a str instance being sent to produce its value (A is parameterized with str). It can't fail (E is parameterized with Never), and it doesn't produce a value (R is parameterized with None).

Never is quite frequently used as the parameter for E, so stateless also supplies a type alias Depend with just that:

from typing import Never

from stateless import Effect


type Depend[A, R] = Effect[A, Never, R]

So hello_world could also have been written:

from stateless import Depend


def hello_world() -> Depend[str, None]:
    message = yield str
    print(message)

To run an Effect, you need an instance of stateless.Runtime. Runtime has just two methods: use and run. Let's look at their definitions:

from stateless import Effect


class Runtime[A]:
    def use[A2](self, ability: A2) -> Runtime[A | A2]:
        ...

    def run[E: Exception, R](self, effect: Effect[A, E, R]) -> R:
        ...

The type parameter A of runtime again stands for "Ability". This is the type of abilities that this Runtime instance can provide.

Runtime.use takes an instance of A, the ability type, to be sent to the effect passed to run upon request (i.e when its type is yielded by the effect).

Runtime.run returns the result of running the Effect (or raises an exception if the effect fails).

Let's run hello_world:

from stateless import Runtime


runtime = Runtime().use(b"Hello, world!")
runtime.run(hello_world())  # type-checker error!

Whoops! We accidentally provided an instance of bytes instead of str, which was required by hello_world. Let's try again:

from stateless import Runtime


runtime = Runtime().use("Hello, world!")
runtime.run(hello_world())  # outputs: Hello, world!

Cool. Okay maybe not. The hello_world example is obviously contrived. There's no real benefit to sending message to hello_world via yield over just providing it as a regular function argument. The example is included here just to give you a rough idea of how the different pieces of stateless fit together.

One thing to note is that the A type parameter of Effect and Runtime work together to ensure type safe dependency injection of abilities: You can't forget to provide an ability (or dependency if you will) to an effect without getting a type error. We'll discuss in more detail later when it makes sense to use abilities for dependency injection, and when it makes sense to use regular function arguments.

Let's look at a bigger example. The main point of a purely functional effect system is to enable side-effects such as IO in a purely functional way. So let's implement some abilities for doing side-effects.

We'll start with an ability we'll call Console for writing to the console:

class Console:
    def print(self, line: str) -> None:
        print(line)

We can use Console with Effect as an ability. Recall that the "send" type of Effect is Any. In order to tell our type checker that the result of yielding the Console class will be a Console instance, we can use the stateless.depend function. Its signature is:

from typing import Type

from stateless import Depend


def depend[A](ability: Type[A]) -> Depend[A, A]:
    ...

So depend just yields the ability type for us, and then returns the instance that will eventually be sent from Runtime.

Let's see that in action with the Console ability:

from stateless import Depend, depend


def say_hello() -> Depend[Console, None]:
    console = yield from depend(Console)
    console.print(f"Hello, world!")

You can of course also just annotate console if you prefer:

from stateless import Depend


def say_hello() -> Depend[Console, None]:
    console: Console = yield Console
    console.print(f"Hello, world!")

Let's add another ability Files to read rom the file system:

class Files:
    def read(self, path: str) -> str:
        with open(path, 'r') as f:
            return f.read()

Putting it all together:

from stateless import Depend


def print_file(path: str) -> Depend[Console | Files, None]:
    ...

Note that A is parameterized with Console | Files since print_file depends on both Console and Files (i.e it will yield both classes).

Let's add a body for print_file:

from stateless import Depend, depend


def print_file(path: str) -> Depend[Console | Files, None]:
    files = yield from depend(Files)
    console = yield from depend(Console)

    content = files.read(path)
    console.print(content)

print_file is a good demonstration of why the "send" type of Effect must be Any: Since print_file expects to be sent instances of Console or File, it's not possible for our type-checker to know on which yield which type is going to be sent, and because of the variance of typing.Generator, we can't write depend in a way that would allow us to type Effect with a "send" type other than Any.

depend is a good example of how you can build complex effects using functions that return simpler effects using yield from:

from stateless import Depend, depend


def get_str() -> Depend[str, str]:
    s = yield from depend(str)
    return s


def get_int() -> Depend[str | int, tuple[str, int]]:
    s = yield from get_str()
    i = yield from depend(int)

    return (s, i)

It will often make sense to use an abc.ABC as your ability type to enforce programming towards the interface and not the implementation. If you use mypy however, note that using abstract classes where typing.Type is expected is a type-error, which will cause problems if you pass an abstract type to depend. We recommend disabling this check, which will also likely be the default for mypy in the future.

you can of course run print_file with Runtime:

from stateless import Runtime


runtime = Runtime().use(Files()).use(Console())
runtime.run(print_file('foo.txt'))

Again, if we forget to supply an ability for runtime required by print_file, we'll get a type error.

Of course the main purpose of dependency injection is to vary the injected ability to change the behavior of the effect. For example, we might want to change the behavior of print_files in tests:

class MockConsole(Console):
    def print(self, line: str) -> None:
        pass


class MockFiles(Files):
    def __init__(self, content: str) -> None:
        self.content = content

    def read(self, path: str) -> str:
        return self.content


console: Console = MockConsole()
files: Files = MockFiles('mock content'.)

runtime = Runtime().use(console).use(files)
runtime.run(print_file('foo.txt'))

Our type-checker will likely infer the types console and files to be MockConsole and MockFiles respectively, so we need to annotate them with the super-types Console and Files. Otherwise it will cause the inferred type of runtime to be Runtime[MockConsole, MockFiles] which would not be type-safe when calling run with an argument of type Effect[Console | Files, Never, None] due to the variance of collections.abc.Generator.

Besides Effect and Depend, stateless provides you with a few other type aliases that can save you a bit of typing. Firstly success which is just defined as:

from typing import Never


type Success[R] = Effect[Never, Never, R]

for effects that don't fail and don't require abilities (can be easily instantiated using the stateless.success function).

Secondly the Try type alias, defined as:

from typing import Never


type Try[E, R] = Effect[Never, E, R]

For effects that do not require abilities, but might fail.

Sometimes, instantiating abilities may itself require side-effects. For example, consider a program that requires a Config ability:

from stateless import Depend


class Config:
    ...


def main() -> Depend[Config, None]:
    ...

Now imagine that you want to provide the Config ability by reading from environment variables:

import os

from stateless import Depend, depend


class OS:
    environ: dict[str, str] = os.environ


def get_config() -> Depend[OS, Config]:
    os = yield from depend(OS)
    return Config(
        url=os.environ['AUTH_TOKEN'],
        auth_token=os.environ['URL']
    )

To supply the Config instance returned from get_config, we can use Runtime.use_effect:

from stateless import Runtime


Runtime().use(OS()).use_effect(get_config()).run(main())

Runtime.use_effect assumes that all abilities required by the effect given as its argument can be provided by the runtime. If this is not the case, you'll get a type-checker error:

from stateless import Depend, Runtime


class A:
    pass


class B:
    pass


def get_B() -> Depend[A, B]:
    ...

Runtime().use(A()).use_effect(get_B())  # OK
Runtime().use_effect(get_B())           # Type-checker error!

Error Handling

So far we haven't used the error type E for anything: We've simply parameterized it with typing.Never. We've claimed that this means that the effect doesn't fail. This is of course not literally true, as exceptions can still occur even if we parameterize E with Never.

Take the Files ability from the previous section for example. Reading from the file system can of course fail for a number of reasons, which in Python will result in a subtype of OSError being raised. So calling for example print_file might raise an exception:

from stateless import Depend


def f() -> Depend[Files, None]:
    yield from print_file('doesnt_exist.txt')  # raises FileNotFoundError

So what's the point of E?

The point is that programming errors can be grouped into two categories: recoverable errors and unrecoverable errors. Recoverable errors are errors that are expected, and that users of the API we are writing might want to know about. FileNotFoundError is an example of such an error.

Unrecoverable errors are errors that there is no point in telling the users of your API about. Depending on the context, ZeroDivisionError or KeyError might be examples of unrecoverable errors.

The intended use of E is to model recoverable errors so that users of your API can handle them with type safety.

Let's use E to model the errors of Files.read_file:

from stateless import Effect, throw


def read_file(path: str) -> Effect[Files, OSError, str]:
    files = yield Files
    try:
        return files.read_file(path)
    except OSError as e:
        return (yield from throw(e))

The signature of stateless.throw is

from typing import Never

from stateless import Effect


def throw[E: Exception](e: E) -> Effect[Never, E, Never]:
    ...

In words throw returns an effect that just yields e and never returns. Because of this signature, if you assign the result of throw to a variable, you have to annotate it. But there is no meaningful type to annotate it with. So you're better off using the somewhat strange looking syntax return (yield from throw(e)).

At a slightly higher level you can use stateless.throws that just catches exceptions and yields them as an effect

from stateless import Depend, throws


@throws(OSError)
def read_file(path: str) -> Depend[Files, str]:
    files = yield Files
    return files.read_file(path)


reveal_type(read_file)  # revealed type is: def (str) -> Effect[Files, OSError, str]

Error handling in stateless is done using the stateless.catch decorator. Its signature is:

from stateless import Effect, Depend


def catch[**P, A, E: Exception, R](
    f: Callable[P, [Effect[A, E, R]]]
) -> Callable[P, Depend[A, E | R]]:
    ...

In words, the catch decorator moves the error from the yield type of the Effect produced by its argument to the return type of the effect of the function returned from catch. This means you can access the potential errors directly in your code:

from stateless import Depend


def handle_errors() -> Depend[Files, str]:
    result: OSError | str = yield from catch(read_file)('foo.txt')
    match result:
        case OSError:
            return 'default value'
        case str():
            return result

(You don't need to annotate the type of result, it can be inferred by your type checker. We do it here simply because its instructive to look at the types.)

Consequently you can use your type checker to avoid unintentionally unhandled errors, or ignore them with type-safety as you please.

catch is a good example of a pattern used in many places in stateless: using decorators to change the result of an effect. The reason for this pattern is that generators are mutable objects.

For example, we could have defined catch like this:

def bad_catch(effect: Effect[A, E, R]) -> Depend[A, E | R]:
    ...

But with this signature, it would not be possible to implement bad_catch without mutating effect as a side-effect, since it's necessary to yield from it to implement catching.

In general, it's not a good idea to write functions that take effects as arguments directly, because it's very easy to accidentally mutate them which would be confusing for the caller:

def f() -> Depend[str, int]:
    ...


def dont_do_this(e: Depend[str, int]) -> Depend[str, int]:
    i = yield from e
    return i


def this_is_confusing() -> Depend[str, tuple[int, int]]:
    e = f()
    r = yield from dont_do_this(e)
    r_again = yield from e  # e was already exhausted, so 'r_again' is None!
    return (r, r_again)

A better idea is to write a decorator that accepts a function that returns effects. That way there is no risk of callers passing generators and then accidentally mutating them as a side effect:

def do_this_instead[**P](f: Callable[P, Depend[str, int]]) -> Callable[P, Depend[str, int]]:
    @wraps(f)
    def decorator(*args: P.args, **kwargs: P.kwargs) -> Depend[str, int]:
        i = yield from f(*args, **kwargs)
        return i
    return decorator


def this_is_easy():
    e = f()
    r = yield from do_this_instead(f)()
    r_again = yield from e
    return (r, r_again)

Parallel Effects

Two challenges present themselves when running generator based effects in parallel:

  • Generators aren't thread-safe.
  • Generators can't be pickled.

Hence, instead of sharing effects between threads and processes to run them in parallel, stateless gives you tools to share functions that return effects plus arguments to those functions between threads and processes.

stateless calls a function that returns an effect plus arguments to pass to that function a task, represented by the stateless.parallel.Task class.

stateless provides two decorators for instantiating Task instances: stateless.parallel.thread and stateless.parallel.process. Their signatures are:

from typing import Callable

from stateless import Effect
from stateless.parallel import Task


def process[**P, A, E: Exception, R](f: Callable[P, Effect[A, E, R]]) -> Callable[P, Task[A, E, R]]:
    ...

def thread[**P, A, E: Exception, R](f: Callable[P, Effect[A, E, R]]) -> Callable[P, Task[A, E, R]]:
    ...

Decorating functions with stateless.parallel.thread indicate to stateless your intention for the resulting task to be run in a separate thread. Decorating functions with stateless.parallel.process indicate your intention for the resulting task to be run in a separate process.

Because of the GIL, using stateless.parallel.thread only makes sense for functions returning effects that are I/O bound. For CPU bound effects, you will want to use stateless.parallel.process.

To run effects in parallel, you use the stateless.parallel function. It's signature is roughly:

from stateless import Effect
from stateless.parallel import Parallel


def parallel[A, E: Exception, R](*tasks: Task[A, E, R]) -> Effect[A | Parallel, E, tuple[R, ...]]:
    ...

(in reality parallel is overloaded to correctly union abilities and errors, and reflect the result types of each effect in the result type of the returned effect.)

In words, parallel accepts a variable number of tasks as its argument, and returns a new effect that depends on the stateless.parallel.Parallel ability. When executed, the effect returned by parallel will run the tasks given as its arguments concurrently.

Here is a full example:

from stateless import parallel, Success, success, Depend
from stateless.parallel import thread, process, Parallel


def sing() -> Success[str]:
    return success("🎡")


def duet() -> Depend[Parallel, tuple[str, str]]:
    result = yield from parallel(
        thread(sing)(),
        process(sing)()
    )
    return result

When using the Parallel ability, you must use it as a context manager, because it manages multiple resources to enable concurrent execution of effects:

from stateless import Runtime
from stateless.parallel import Parallel


with Parallel() as ability:
    print(Runtime().use(ability).run(duet()))  # outputs: ("🎡", "🎡")

In this example the first sing invocation will be run in a separate thread because its wrapped with thread, and the second sing invocation will be run in a separate process because it's wrapped by process. Note that although thread and process are strictly speaking decorators, they don't return stateless.Effect instances. For this reason, it's probably not a good idea to use them as @thread or @process, since this reduces the re-usability of the decorated function. Use them at the call site as shown in the example instead.

stateless.parallel.Task does however implement __iter__ to return the result of the decorated function, so you can yield from them if necessary:

from stateless import Success, thread


def sing_more() -> Success[str]:
    # This is rather pointless,
    # but helps you out if you for some
    # reason have used @thread instead of thread(...)
    note = yield from thread(sing)()
    return note * 2

If you need more control over the resources managed by stateless.parallel.Parallel, you can pass them as arguments:

from multiprocessing.pool import ThreadPool
from multiprocessing import Manager

from stateless.parallel import Parallel


with (
    Manager() as manager,
    manager.Pool() as pool,
    ThreadPool() as thread_pool,
    Parallel(thread_pool, pool) as parallel
):
    ...

The process pool used to execute stateless.parallel.Task instances needs to be run with a manager because it needs to be sent to the process executing the task in case it needs to run more effects in other processes.

Note that if you pass in in the thread pool and proxy pool as arguments, stateless.parallel.Parallel will not exit them for you when it itself exits: you need to manage their state yourself.

You can of course subclass stateless.parallel.Parallel to change the interpretation of this ability (for example in tests). The two main functions you'll want to override is run_thread_tasks and run_cpu_tasks:

from stateless import Runtime, Effect
from stateless.parallel import Parallel, Task


class MockParallel(Parallel):
    def __init__(self):
        pass

    def run_cpu_tasks(self,
                 runtime: Runtime[object],
                 tasks: Sequence[Task[object, Exception, object]]) -> Tuple[object, ...]:
        return tuple(runtime.run(iter(task)) for task in tasks)

    def run_thread_tasks(self
                    runtime: Runtime[object],
                    effects: Sequence[Effect[object, Exception, object]]) -> Tuple[object, ...]:
        return tuple(runtime.run(iter(task)) for task in tasks)

Repeating and Retrying Effects

A stateless.Schedule is a type with an __iter__ method that returns an effect producing an iterator of timedelta instances. It's defined like:

from typing import Protocol, Iterator
from datetime import timedelta

from stateless import Depend


class Schedule[A](Protocol):
    def __iter__(self) -> Depend[A, Iterator[timedelta]]:
        ...

The type parameter A is present because some schedules may require abilities to complete.

The stateless.schedule module contains a number of of helpful implemenations of Schedule, for example Spaced or Recurs.

Schedules can be used with the repeat decorator, which takes schedule as its first argument and repeats the decorated function returning an effect until the schedule is exhausted or an error occurs:

from datetime import timedelta

from stateless import repeat, success, Success, Runtime
from stateless.schedule import Recurs, Spaced
from stateless.time import Time


@repeat(Recurs(2, Spaced(timedelta(seconds=2))))
def f() -> Success[str]:
    return success("hi!")


print(Runtime().use(Time()).run(f()))  # outputs: ("hi!", "hi!")

Effects created through repeat depends on the Time ability from stateless.time because it needs to sleep between each execution of the effect.

Schedules are a good example of a pattern used a lot in stateless: Classes with an __iter__ method that returns effects.

This is a useful pattern because such objects can be yielded from in functions returning effects multiple times where a new generator will be instantiated every time:

def this_works() -> Success[timedelta]:
    schedule = Spaced(timedelta(seconds=2))
    deltas = yield from schedule
    deltas_again = yield from schedule  # safe!
    return deltas

For example, repeat needs to yield from the schedule given as its argument to repeat the decorated function. If the schedule was just a generator it would only be possible to yield from the schedule the first time f in this example was called.

stateless.retry is like repeat, except that it returns succesfully when the decorated function yields no errors, or fails when the schedule is exhausted:

from datetime import timedelta

from stateless import retry, throw, Try, throw, success, Runtime
from stateless.schedule import Recurs, Spaced
from stateless.time import Time


fail = True


@retry(Recurs(2, Spaced(timedelta(seconds=2))))
def f() -> Try[RuntimeError, str]:
    global fail
    if fail:
        fail = False
        return throw(RuntimeError('Whoops...'))
    else:
        return success('Hooray!')


print(Runtime().use(Time()).run(f()))  # outputs: 'Hooray!'

Memoization

Effects can be memoized using the stateless.memoize decorator:

from stateless import memoize, Depend
from stateless.console import Console, print_line


@memoize
def f() -> Depend[Console, str]:
    yield from print_line('f was called')
    return 'done'


def g() -> Depend[Console, tuple[str, str]]:
    first = yield from f()
    second = yield from f()
    return first, second


result = Runtime().use(Console()).run(f())  # outputs: 'f was called' once, even though the effect was yielded twice

print(result)  # outputs: ('done', 'done')

memoize works like functools.lru_cache, in that the memoized effect is cached based on the arguments of the decorated function. In fact, memoize takes the same parameters as functools.lru_cache (maxsize and typed) with the same meaning.

Known Issues

See the issues page.

Algebraic Effects Vs. Monads

All functional effect system work essentially the same way:

  1. Programs send a description of the side-effect needed to be performed to the effect system and pause their executing while the effect system handles the side-effect.
  2. Once the result of performing the side-effect is ready, program execution is resumed at the point it was paused

Step 2. is the tricky part: how can program execution be resumed at the point it was paused?

Monads are the most common solution. When programming with monads, in addition to supplying the effect system with a description of a side-effect, the programmer also supplies a function to be called with the result of handling the described effect. In functional programming such a function is called a continuation. In other paradigms it might be called a callback function.

For example it might look like this:

def say_hello() -> IO[None]:
    return Input("whats your name?").bind(lambda name: Print(f"Hello, {name}!"))

One of the main benefits of basing effect systems on monads is that they don't rely on any special language features: its all literally just functions.

However, many programmers find monads awkward. Programming with callback functions often lead to code thats hard for humans to parse, which has ultimately inspired specialized language features for hiding the callback functions with syntax sugar like Haskell's do notation, or for comprehensions in Scala.

Moreover, monads famously do not compose, meaning that when writing code that needs to juggle multiple types of side-effects (like errors and IO), it's up to the programmer to pack and unpack results of various types of effects (or use advanced features like monad transformers which come with their own set of problems).

Additionally, in languages with dynamic binding such as Python, calling functions is relatively expensive, which means that using callbacks as the principal method for resuming computation comes with a fair amount of performance overhead.

Finally, interpreting monads is often a recursive procedure, meaning that it's necessary to worry about stack safety in languages without tail call optimisation such as Python. This is usually solved using trampolines which further adds to the performance overhead.

Because of all these practical challenges of programming with monads, people have been looking for alternatives. Algebraic effects is one the things suggested that address many of the challenges of monadic effect systems.

In algebraic effect systems, such as stateless, the programmer still supplies the effect system with a description of the side-effect to be carried out, but instead of supplying a callback function to resume the computation with, the result of handling the effect is returned to the point in program execution that the effect description was produced. The main drawback of this approach is that it requires special language features to do this. In Python however, such a language feature does exist: Generators and coroutines.

Using coroutines for algebraic effects solves many of the challenges with monadic effect systems:

  • No callback functions are required, so readability and understandability of the effectful code is much more straightforward.
  • Code that needs to describe side-effects can simply list all the effects it requires, so there is no composition problem.
  • There are no callback functions, so no need to worry about performance overhead of calling a large number of functions or using trampolines to ensure stack safety.

Background

Similar Projects