amphp/amp

AMPHP is a collection of event-driven libraries for PHP designed with fibers and concurrency in mind. amphp/amp specifically provides futures and cancellations as fundamental primitives for asynchronous programming. We're now using Revolt instead of shipping an event loop implementation with amphp/amp.

Amp makes heavy use of fibers shipped with PHP 8.1 to write asynchronous code just like synchronous, blocking code. In contrast to earlier versions, there's no need for generator based coroutines or callbacks. Similar to threads, each fiber has its own call stack, but fibers are scheduled cooperatively by the event loop. Use Amp\async() to run things concurrently.

Motivation

Traditionally, PHP follows a sequential execution model. The PHP engine executes one line after the other in sequential order. Often, however, programs consist of multiple independent sub-programs which can be executed concurrently.

If you query a database, you send the query and wait for the response from the database server in a blocking manner. Once you have the response, you can start doing the next thing. Instead of sitting there and doing nothing while waiting, we could already send the next database query, or do an HTTP call to an API. Let's make use of the time we usually spend on waiting for I/O!

Revolt allows such concurrent I/O operations. We keep the cognitive load low by avoiding callbacks. Our APIs can be used like any other library, except that things also work concurrently, because we use non-blocking I/O under the hood. Run things concurrently using Amp\async() and await the result using Future::await() where and when you need it!

There have been various techniques for implementing concurrency in PHP over the years, e.g. callbacks and generators shipped in PHP 5. These approaches suffered from the "What color is your function" problem, which we solved by shipping Fibers with PHP 8.1. They allow for concurrency with multiple independent call stacks.

Fibers are cooperatively scheduled by the event-loop, which is why they're also called coroutines. It's important to understand that only one coroutine is running at any given time, all other coroutines are suspended in the meantime.

You can compare coroutines to a computer running multiple programs using a single CPU core. Each program gets a timeslot to execute. Coroutines, however, are not preemptive. They don't get their fixed timeslot. They have to voluntarily give up control to the event loop.

Any blocking I/O function blocks the entire process while waiting for I/O. You'll want to avoid them. If you haven't read the installation guide, have a look at the Hello World example that demonstrates the effect of blocking functions. The libraries provided by AMPHP avoid blocking for I/O.

Installation

This package can be installed as a Composer dependency.

composer require amphp/amp

If you use this library, it's very likely you want to schedule events using Revolt, which you should require separately, even if it's automatically installed as a dependency.

composer require revolt/event-loop

These packages provide the basic building blocks for asynchronous / concurrent applications in PHP. We offer a lot of packages building on top of these, e.g.

• amphp/byte-stream providing a stream abstraction
• amphp/socket providing a socket layer for UDP and TCP including TLS
• amphp/parallel providing parallel processing to utilize multiple CPU cores and offload blocking operations
• amphp/http-client providing an HTTP/1.1 and HTTP/2 client
• amphp/http-server providing an HTTP/1.1 and HTTP/2 application server
• amphp/mysql and amphp/postgres for non-blocking database access
• and many more packages

Requirements

This package requires PHP 8.1 or later. No extensions required!

Extensions are only needed if your app necessitates a high numbers of concurrent socket connections, usually this limit is configured up to 1024 file descriptors.

Usage

Coroutines

Coroutines are interruptible functions. In PHP, they can be implemented using fibers.

Note Previous versions of Amp used generators for a similar purpose, but fibers can be interrupted anywhere in the call stack making previous boilerplate like Amp\call() unnecessary.

At any given time, only one fiber is running. When a coroutine suspends, execution of the coroutine is temporarily interrupted, allowing other tasks to be run. Execution is resumed once a timer expires, stream operations are possible, or any awaited Future completes.

Low-level suspension and resumption of coroutines is handled by Revolt's Suspension API.

<?php

require __DIR__ . '/vendor/autoload.php';

use Revolt\EventLoop;

$suspension = EventLoop::getSuspension();

EventLoop::delay(5, function () use ($suspension): void {
    print '++ Executing callback created by EventLoop::delay()' . PHP_EOL;

    $suspension->resume(null);
});

print '++ Suspending to event loop...' . PHP_EOL;

$suspension->suspend();

print '++ Script end' . PHP_EOL;

Callbacks registered on the Revolt event-loop are automatically run as coroutines and it's safe to suspend them. Apart from the event-loop API, Amp\async() can be used to start an independent call stack.

<?php

use function Amp\delay;

require __DIR__ . '/vendor/autoload.php';

Amp\async(function () {
    print '++ Executing callback passed to async()' . PHP_EOL;

    delay(3);

    print '++ Finished callback passed to async()' . PHP_EOL;
});

print '++ Suspending to event loop...' . PHP_EOL;
delay(5);

print '++ Script end' . PHP_EOL;

Future

A Future is an object representing the eventual result of an asynchronous operation. There are three states:

• Completed successfully: The future has been completed successfully.
• Errored: The future failed with an exception.
• Pending: The future is still pending.

A successfully completed future is analog to a return value, while an errored future is analog to throwing an exception.

One way to approach asynchronous APIs is using callbacks that are passed when the operation is started and called once it completes:

doSomething(function ($error, $value) {
    if ($error) {
        /* ... */
    } else {
        /* ... */
    }
});

The callback approach has several drawbacks.

• Passing callbacks and doing further actions in them that depend on the result of the first action gets messy really quickly.
• An explicit callback is required as input parameter to the function, and the return value is simply unused. There's no way to use this API without involving a callback.

That's where futures come into play. They're placeholders for the result that are returned like any other return value. The caller has the choice of awaiting the result using Future::await() or registering one or several callbacks.

try {
    $value = doSomething()->await();
} catch (...) {
    /* ... */
}

Combinators

In concurrent applications, there will be multiple futures, where you might want to await them all or just the first one.

await

Amp\Future\await($iterable, $cancellation) awaits all Future objects of an iterable. If one of the Future instances errors, the operation will be aborted with that exception. Otherwise, the result is an array matching keys from the input iterable to their completion values.

The await() combinator is extremely powerful because it allows you to concurrently execute many asynchronous operations at the same time. Let's look at an example using amphp/http-client to retrieve multiple HTTP resources concurrently:

<?php

use Amp\Future;
use Amp\Http\Client\HttpClientBuilder;
use Amp\Http\Client\Request;

$httpClient = HttpClientBuilder::buildDefault();
$uris = [
    "google" => "https://www.google.com",
    "news"   => "https://news.google.com",
    "bing"   => "https://www.bing.com",
    "yahoo"  => "https://www.yahoo.com",
];

try {
    $responses = Future\await(array_map(function ($uri) use ($httpClient) {
        return Amp\async(fn () => $httpClient->request(new Request($uri, 'HEAD')));
    }, $uris));

    foreach ($responses as $key => $response) {
        printf(
            "%s | HTTP/%s %d %s\n",
            $key,
            $response->getProtocolVersion(),
            $response->getStatus(),
            $response->getReason()
        );
    }
} catch (Exception $e) {
    // If any one of the requests fails the combo will fail
    echo $e->getMessage(), "\n";
}

awaitAnyN

Amp\Future\awaitAnyN($count, $iterable, $cancellation) is the same as await() except that it tolerates individual errors. A result is returned once exactly $count instances in the iterable complete successfully. The return value is an array of values. The individual keys in the component array are preserved from the iterable passed to the function for evaluation.

awaitAll

Amp\Promise\awaitAll($iterable, $cancellation) awaits all futures and returns their results as [$errors, $values] array.

awaitFirst

Amp\Promise\awaitFirst($iterable, $cancellation) unwraps the first completed Future, whether successfully completed or errored.

awaitAny

Amp\Promise\awaitAny($iterable, $cancellation) unwraps the first successfully completed Future.

Future Creation

Futures can be created in several ways. Most code will use Amp\async() which takes a function and runs it as coroutine in another Fiber.

Sometimes an interface mandates a Future to be returned, but results are immediately available, e.g. because they're cached. In these cases Future::complete(mixed) and Future::error(Throwable) can be used to construct an immediately completed Future.

DeferredFuture

Note The DeferredFuture API described below is an advanced API that many applications probably don't need. Use Amp\async() or combinators instead where possible.

Amp\DeferredFuture is responsible for completing a pending Future. You create a Amp\DeferredFuture and uses its getFuture method to return an Amp\Future to the caller. Once result is ready, you complete the Future held by the caller using complete or error on the linked DeferredFuture.

final class DeferredFuture
{
    public function getFuture(): Future;
    public function complete(mixed $value = null);
    public function error(Throwable $throwable);
}

Warning If you're passing DeferredFuture objects around, you're probably doing something wrong. They're supposed to be internal state of your operation.

Warning You can't complete a future with another future; Use Future::await() before calling DeferredFuture::complete() in such cases.

Here's a simple example of an asynchronous value producer asyncMultiply() creating a DeferredFuture and returning the associated Future to its caller.

<?php // Example async producer using DeferredFuture

use Revolt\EventLoop;

function asyncMultiply(int $x, int $y): Future
{
    $deferred = new Amp\DeferredFuture;

    // Complete the async result one second from now
    EventLoop::delay(1, function () use ($deferred, $x, $y) {
        $deferred->complete($x * $y);
    });

    return $deferred->getFuture();
}

$future = asyncMultiply(6, 7);
$result = $future->await();

var_dump($result); // int(42)

Cancellation

Every operation that supports cancellation accepts an instance of Cancellation as argument. Cancellations are objects that allow registering handlers to subscribe to cancellation requests. These objects are passed down to sub-operations or have to be handled by the operation itself.

$cancellation->throwIfRequested() can be used to fail the current operation with a CancelledException once cancellation has been requested. While throwIfRequested() works well, some operations might want to subscribe with a callback instead. They can do so using Cancellation::subscribe() to subscribe any cancellation requests that might happen.

The caller creates a Cancellation by using one of the implementations below.

Note Cancellations are advisory only. A DNS resolver might ignore cancellation requests after the query has been sent as the response has to be processed anyway and can still be cached. An HTTP client might continue a nearly finished HTTP request to reuse the connection, but might abort a chunked encoding response as it cannot know whether continuing is actually cheaper than aborting.

TimeoutCancellation

A TimeoutCancellations automatically cancels itself after the specified number of seconds.

request("...", new Amp\TimeoutCancellation(30));

SignalCancellation

A SignalCancellation automatically cancels itself after a specified signal has been received by the current process.

request("...", new Amp\SignalCancellation(SIGINT));

DeferredCancellation

A DeferredCancellation allows manual cancellation with the call of a method. This is the preferred way if you need to register some custom callback somewhere instead of shipping your own implementation. Only the caller has access to the DeferredCancellation and can cancel the operation using DeferredCancellation::cancel().

$deferredCancellation = new Amp\DeferredCancellation();

// Register some custom callback somewhere
onSomeEvent(fn () => $deferredCancellation->cancel());

request("...", $deferredCancellation->getCancellation());

NullCancellation

A NullCancellation will never be cancelled. Cancellation is often optional, which is usually implemented by making the parameter nullable. To avoid guards like if ($cancellation), a NullCancellation can be used instead.

$cancellation ??= new NullCancellationToken();

CompositeCancellation

A CompositeCancellation combines multiple independent cancellation objects. If any of these cancellations is cancelled, the CompositeCancellation itself will be cancelled.

Versioning

amphp/amp follows the semver semantic versioning specification like all other amphp packages.

Compatible Packages

Compatible packages should use the amphp topic on GitHub.

Security

If you discover any security related issues, please email me@kelunik.com instead of using the issue tracker.

License

The MIT License (MIT). Please see LICENSE for more information.