#sender #sender-receiver #receivers #execution

senders_receivers

Senders/Receivers implementation for Rust

7 releases

0.0.7 Aug 10, 2024
0.0.6 Aug 8, 2024
0.0.4 Jul 25, 2024

#349 in Asynchronous

Download history 45/week @ 2024-07-27 227/week @ 2024-08-03 123/week @ 2024-08-10 4/week @ 2024-08-17 12/week @ 2024-09-14 1/week @ 2024-09-21 29/week @ 2024-09-28 1/week @ 2024-10-05

358 downloads per month

BSD-2-Clause

415KB
11K SLoC

This is an implementation (for certain value of implementation) of C++ senders/receivers.

Please refer to the rust documentation for how to use this.

Note

This code is very much a work-in-progress, and large parts are still missing. The API is also still in flux.

A Tiny Example

use senders_receivers::*;

let sender = Just::from((1, 2, 3, 4))
           | Then::from(|(a, b, c, d)| (a * b * c * d,));
println!("outcome: {}", sender.sync_wait().expect("no error").expect("no cancelation").0);

What this does:

  • Just::from: declares an starting value for the sender chain.
  • Then::from: declares a transformation on these values.
  • | (the pipe symbol): used to bind them together.

None of the steps are run, until sync_wait is invoked.

Senders? Receivers?

The main concept of this code comes from having both senders and receivers. A sender is a type that produces a value signal (or an error signal, or a done signal). A receiver is a type that accepts this signal.

Senders are composable: they can be chained together, creating a new sender. This is done with the | (pipe) symbol.

If you are using the library, you will not spot any receivers, unless you're implementing your own extensions.

Signals

Each sender, produces either a value signal, an error signal, or a done signal. Exactly one of these will be produced.

A value signal indicates that the sender completed successfully, and produced a value.
An error signal indicates that the sender failed, and produced an [Error].
A done signal indicates that the sender canceled its work (producing neither a value, nor an error).

Schedulers

Each operation will run on a [Scheduler]. (Sometimes more than one, for example if you [Transfer] to a different scheduler.)

A scheduler encapsulates the concept of CPU-time. Different schedulers will have different characteristics, related to when/where things run.

Currently, the following schedulers are implemented:

  • [ImmediateScheduler] runs every task immediately. This is more-or-less the default scheduler.
  • ThreadPool runs tasks using a threadpool.

Comparison With C++

An attempt is made to remain close the the usability of its C++ counterpart, but sacrifices had to be made.

  • The Error type is type-erased/dynamic.
    For a non-dynamic error type, we would require a variant-type with variadic type arguments.
  • The Value type is a tuple.
    For a non-tuple type, we would require variadic type arguments.
  • The Value is not a variant.
    For a variant-type, we would require a variant-type with variadic type arguments. Furthermore, rust closures bind their type on first use, dispatching to overloads would be difficult.
  • Scheduler-based specializations are omitted.
    I think for this, we would require specializations. (Although the io::write implementation does permit specialization, using an opt-in for a default implementation.)
  • Connect can no longer fail.
    In C++, the connect call can fail, and this will result in an error being propagated. This meant that when the connect call inside a let_value step fails, the error would have to propage via the receiver. The same receiver would also be passed to the connect call, using move semantics.
    In rust, this will cause the borrow checker to flag this as bad, and I kinda like the borrow checker. Instead I decided: if the connect fails, it'll have to create an operation state that'll propagate an error.
  • Operation-states don't nest themselves.
    Mostly, the constraint stems from rust being very picky about move semantics (and I approve). And also, rust does not allow in-place construction (which is required for nested operation states; the C++ design requires in-place construction). Instead we wrap subsequent operations into a receiver, during the connect call. This has the small disadvantage we cannot do any preparations during the OperationState::start step.
  • [LetValue] arguments are passed by reference (similar to C++) but cannot be taken ownership of.
    Instead, the arguments are retained and combined with the values from the sender-chain returned by the LetValue function.

Some sacrifices stem from me disagreeing with the C++ design: I liked the promise from the design, that scheduler changes are explicit only. But in practice, it was very hard to use, and my code, once async, always needed to grab the receiver-scheduler (because it was too common for the sender not to have a (known) scheduler).

  • The done and error channels no longer have an associated scheduler.
    There is no way to properly guarantee which scheduler an error is propagated upon.
  • The value channel now always has an associated scheduler.
    This allows for example a non-blocking write (ex: aio_write) to suspend execution, and resume on the same scheduler code was running on earlier.
  • The receiver no longer has a scheduler.
    Since the value signal now has a scheduler, we no longer need one. (Also, the C++ receivers have an optional scheduler, which makes coding against them really cumbersome.)
  • Schedulers have a LocalScheduler type.
    For most schedulers, that will be the scheduler itself. This allows for schedulers to declare follow-up code to run on a different scheduler. For example, the embarrasingly-parallel scheduler schedules on any thread the first time, and then propagates a scheduler that'll always use the same thread for subsequent scheduling.
  • Error/done recovery operations must complete with the same scheduler-type and value-type, as what would be sent during the value path.
    This is a consequence of us limiting things to a single type, and sending the scheduler along in the value-signal. (Note that the error and done signals don't propagate a scheduler alongside; this would be impossible without making [LetValue] a lot more cumbersome to use.)

The reason that error channels no longer have an associated scheduler, is because scheduler-transfers can fail, and this would break the invariant of an error-scheduler.

Dropping the scheduler from done and error signals, means that scheduler switches will only happen on the happy path (and on recovery paths).

Dependencies

~0.8–11MB
~65K SLoC