NAME

Future::AsyncAwait - deferred subroutine syntax for futures

SYNOPSIS

use v5.14;
use Future::AsyncAwait;

async sub do_a_thing
{
   my $first = await do_first_thing();

   my $second = await do_second_thing();

   return combine_things( $first, $second );
}

do_a_thing()->get;

DESCRIPTION

This module provides syntax for deferring and resuming subroutines while waiting for Futures to complete. This syntax aims to make code that performs asynchronous operations using futures look neater and more expressive than simply using then chaining and other techniques on the futures themselves. It is also a similar syntax used by a number of other languages; notably C# 5, EcmaScript 6, Python 3, Dart. Rust is considering adding it.

This module is still under active development. While it now seems relatively stable enough for most use-cases and has received a lot of "battle-testing" in a wide variety of scenarios, there may still be the occasional case of memory leak left in it, especially if still-pending futures are abandoned.

The new syntax takes the form of two new keywords, async and await.

async

The async keyword should appear just before the sub keyword that declares a new function. When present, this marks that the function performs its work in a potentially asynchronous fashion. This has two effects: it permits the body of the function to use the await expression, and it wraps the return value of the function in a Future instance.

async sub myfunc
{
   return 123;
}

my $f = myfunc();
my $result = $f->get;

As well as named function declarations it is also supported on anonymous function expressions.

my $code = async sub { return 456 };
my $f = $code->();
my $result = $f->get;

This async-declared function always returns a Future instance when invoked. The returned future instance will eventually complete when the function returns, either by the return keyword or by falling off the end; the result of the future will be the return value from the function's code. Alternatively, if the function body throws an exception, this will cause the returned future to fail.

If the final expression in the body of the function returns a Future, don't forget to await it rather than simply returning it as it is, or else this return value will become double-wrapped - almost certainly not what you wanted.

async sub otherfunc { ... }

async sub myfunc
{
   ...
   return await otherfunc();
}

await

The await keyword forms an expression which takes a Future instance as an operand and yields the eventual result of it. Superficially it can be thought of similar to invoking the get method on the future.

my $result = await $f;

my $result = $f->get;

However, the key difference (and indeed the entire reason for being a new syntax keyword) is the behaviour when the future is still pending and is not yet complete. Whereas the simple get method would block until the future is complete, the await keyword causes its entire containing function to become suspended, making it return a new (pending) future instance. It waits in this state until the future it was waiting on completes, at which point it wakes up and resumes execution from the point of the await expression. When the now-resumed function eventually finishes (either by returning a value or throwing an exception), this value is set as the result of the future it had returned earlier.

await provides scalar context to its controlling expression.

async sub func {
   # this function is invoked in scalar context
}

await func();

Because the await keyword may cause its containing function to suspend early, returning a pending future instance, it is only allowed inside async-marked subs.

The converse is not true; just because a function is marked as async does not require it to make use of the await expression. It is still useful to turn the result of that function into a future, entirely without awaiting on any itself.

Any function that doesn't actually await anything, and just returns immediate futures can be neatened by this module too.

Instead of writing

sub imm
{
   ...
   return Future->done( @result );
}

you can now simply write

async sub imm
{
   ...
   return @result;
}

with the added side-benefit that any exceptions thrown by the elided code will be turned into an immediate-failed Future rather than making the call itself propagate the exception, which is usually what you wanted when dealing with futures.

await (toplevel)

Since version 0.47.

An await expression is also permitted directly in the main script at toplevel, outside of async sub. This is implemented by simply invoking the get method on the future value. Thus, the following two lines are directly equivalent:

await afunc();
afunc()->get;

This is provided as a syntax convenience for unit tests, toplevel scripts, and so on. It allows code to be written in a style that can be easily moved into an async sub, and avoids encouraging "bad habits" of invoking the get method directly.

CANCEL

Experimental. Since version 0.44.

The CANCEL keyword declares a block of code which will be run in the event that the future returned by the async sub is cancelled.

async sub f
{
   CANCEL { warn "This task was cancelled"; }

   await ...
}

f()->cancel;

A CANCEL block is a self-contained syntax element, similar to perl constructions like BEGIN, and does not need a terminating semicolon.

When a CANCEL block is encountered during execution of the async sub, the code in its block is stored for the case that the returned future is cancelled. Each will take effect as it is executed, possibly multiple times if it appears inside a loop, or not at all if it appears conditionally in a branch that was not executed.

async sub g
{
   if(0) {
      CANCEL { warn "This does not happen"; }
   }

   foreach my $x ( 1..3 ) {
      CANCEL { warn "This happens for x=$x"; }
   }

   await ...
}

g()->cancel;

CANCEL blocks are only invoked if a still-pending future is cancelled. They are discarded without being executed if the function finishes; either successfully or if it throws an exception.

Experimental Features

Some of the features of this module are currently marked as experimental. They will provoke warnings in the experimental category, unless silenced.

You can silence this with no warnings 'experimental' but then that will silence every experimental warning, which may hide others unintentionally. For a more fine-grained approach you can instead use the import line for this module to only silence this module's warnings selectively:

use Future::AsyncAwait qw( :experimental(cancel) );

use Future::AsyncAwait qw( :experimental );  # all of the above

SUPPORTED USES

Most cases involving awaiting on still-pending futures should work fine:

async sub foo
{
   my ( $f ) = @_;

   BEFORE();
   await $f;
   AFTER();
}

async sub bar
{
   my ( $f ) = @_;

   return 1 + await( $f ) + 3;
}

async sub splot
{
   while( COND ) {
      await func();
   }
}

async sub wibble
{
   if( COND ) {
      await func();
   }
}

async sub wobble
{
   foreach my $var ( THINGs ) {
      await func();
   }
}

async sub quux
{
   my $x = do {
      await func();
   };
}

async sub splat
{
   eval {
      await func();
   };
}

Plain lexical variables are preserved across an await deferral:

async sub quux
{
   my $message = "Hello, world\n";
   await func();
   print $message;
}

On perl versions 5.26 and later async sub syntax supports the signatures feature if it is enabled:

use v5.26;
use feature 'signatures';

async sub quart($x, $y)
{
   ...
}

Cancellation

Cancelled futures cause a suspended async sub to simply stop running.

async sub fizz
{
   await func();
   say "This is never reached";
}

my $f = fizz();
$f->cancel;

Cancellation requests can propagate backwards into the future the async sub is currently waiting on.

async sub floof
{
   ...
   await $f1;
}

my $f2 = floof();

$f2->cancel;  # $f1 will be cancelled too

This behaviour is still more experimental than the rest of the logic. The following should be noted:

  • There is currently no way to perform the equivalent of "on_cancel" in Future to add a cancellation callback to a future chain.

  • Cancellation propagation is only implemented on Perl version 5.24 and above. An async sub in an earlier perl version will still stop executing if cancelled, but will not propagate the request backwards into the future that the async sub is currently waiting on. See "TODO".

SUBCLASSING Future

By default when an async sub returns a result or fails immediately before awaiting, it will return a new completed instance of the Future class. In order to allow code that wishes to use a different class to represent futures the module import method can be passed the name of a class to use instead.

use Future::AsyncAwait future_class => "Subclass::Of::Future";

async sub func { ... }

This has the usual lexically-scoped effect, applying only to async subs defined within the block; others are unaffected.

use Future::AsyncAwait;

{
   use Future::AsyncAwait future_class => "Different::Future";
   async sub x { ... }
}

async sub y { ... }  # returns a regular Future

This will only affect immediate results. If the await keyword has to suspend the function and create a new pending future, it will do this by using the prototype constructor on the future it itself is waiting on, and the usual subclass-respecting semantics of "new" in Future will remain in effect there. As such it is not usually necessary to use this feature just for wrapping event system modules or other similar situations.

Such an alternative subclass should implement the API documented by Future::AsyncAwait::Awaitable.

WITH OTHER MODULES

Syntax::Keyword::Try

As of Future::AsyncAwait version 0.10 and Syntax::Keyword::Try version 0.07, cross-module integration tests assert that basic try/catch blocks inside an async sub work correctly, including those that attempt to return from inside try.

use Future::AsyncAwait;
use Syntax::Keyword::Try;

async sub attempt
{
   try {
      await func();
      return "success";
   }
   catch {
      return "failed";
   }
}

Syntax::Keyword::Dynamically

As of Future::AsyncAwait version 0.32, cross-module integration tests assert that the dynamically correctly works across an await boundary.

use Future::AsyncAwait;
use Syntax::Keyword::Dynamically;

our $var;

async sub trial
{
   dynamically $var = "value";

   await func();

   say "Var is still $var";
}

Object::Pad

As of Future::AsyncAwait version 0.38 and Object::Pad version 0.15, both modules now use XS::Parse::Sublike to parse blocks of code. Because of this the two modules can operate together and allow class methods to be written as async subs which await expressions:

use Future::AsyncAwait;
use Object::Pad;

class Example
{
   async method perform($block)
   {
      say "$self is performing code";
      await $block->();
      say "code finished";
   }
}

SEE ALSO

TODO

  • Suspend and resume with some consideration for the savestack; i.e. the area used to implement local and similar. While in general local support has awkward questions about semantics, there are certain situations and cases where internally-implied localisation of variables would still be useful and can be supported without the semantic ambiguities of generic local.

    our $DEBUG = 0;
    
    async sub quark
    {
       local $DEBUG = 1;
       await func();
    }

    Since foreach loops on non-lexical iterator variables (usually the $_ global variable) effectively imply a local-like behaviour, these are also disallowed.

    async sub splurt
    {
       foreach ( LIST ) {
          await ...
       }
    }

    Some notes on what makes the problem hard can be found at

    https://rt.cpan.org/Ticket/Display.html?id=122793

  • Currently this module requires perl version 5.16 or later. Additionally, threaded builds of perl earlier than 5.22 are not supported.

    https://rt.cpan.org/Ticket/Display.html?id=122252

    https://rt.cpan.org/Ticket/Display.html?id=124351

  • Implement cancel back-propagation for Perl versions earlier than 5.24. Currently this does not work due to some as-yet-unknown effects that installing the back-propagation has, causing future instances to be reclaimed too early.

    https://rt.cpan.org/Ticket/Display.html?id=129202

KNOWN BUGS

This is not a complete list of all known issues, but rather a summary of the most notable ones that currently prevent the module from working correctly in a variety of situations. For a complete list of known bugs, see the RT queue at https://rt.cpan.org/Dist/Display.html?Name=Future-AsyncAwait.

  • await inside map or grep blocks does not work. This is due to the difficulty of detecting the map or grep context from internal perl state at suspend time, sufficient to be able to restore it again when resuming.

    https://rt.cpan.org/Ticket/Display.html?id=129748

    As a workaround, consider converting a map expression to the equivalent form using push onto an accumulator array with a foreach loop:

    my @results = map { await func($_) } ITEMS;

    becomes

    my @results;
    foreach my $item ( ITEMS ) {
       push @results, await func($item);
    }

    with a similar transformation for grep expressions.

    Alternatively, consider using the fmap* family of functions from Future::Utils to provide a concurrent version of the same code, which can keep multiple items running concurrently:

    use Future::Utils qw( fmap );
    
    my @results = await fmap { func( shift ) }
       foreach    => [ ITEMS ],
       concurrent => 5;

ACKNOWLEDGEMENTS

With thanks to Zefram, ilmari and others from irc.perl.org/#p5p for assisting with trickier bits of XS logic.

Thanks to genio for project management and actually reminding me to write some code.

Thanks to The Perl Foundation for sponsoring me to continue working on the implementation.

AUTHOR

Paul Evans <leonerd@leonerd.org.uk>