[swift-evolution] [Concurrency] A slightly different perspective

Jonathan Hull jhull at gbis.com
Sun Sep 3 23:15:08 CDT 2017

> On Sep 3, 2017, at 7:35 PM, BJ Homer <bjhomer at gmail.com> wrote:
> Jonathan,
> You've mentioned the desire to have 'async' defer calling 'await', but I haven't seen a detailed design yet.
Oh, we discussed it awhile back on a few other threads.  I am happy to help write up a more formal/detailed design if there is enough interest...

> For example, is the following code valid?
>   let image = async fetchImage()
>   let image2 = async fetchImage()
>   let deferredThings = [image1, image2]
> If so, what is the type of 'deferredThings'? And how does it not count as 'using' the values.

No this code is not valid.  You would need to ‘await’ both image 1 & 2 before they could be put in an array or transferred to another variable. You could combine the ‘await’s though (similar to try):

	let image = async fetchImage()
	let image2 = async fetchImage()
	let deferredThings = await [image1, image2]

Note: You can return something which is deferred from an async function without awaiting though...

> If the above code is not valid, how is this situation better than the suggested use of a Future type to allow concurrent async requests?
>   let future1 = Future { await fetchImage() }
>   let future2 = Future { await fetchImage() }
>   let deferredThings = [future1, future2]
> Note that in this example, 'deferredThings' has a concrete type, and we can inspect its values.

It isn’t meant to be used instead of Futures (though you may not need to reach for them as often), it is a much lower-level construct which would be used as a building block for things like futures (and other concurrency constructs).

Because there is no way for the programmer to get at the thing being ‘await’ed (or any representation of it) without awaiting on it, and it can’t escape the context, it gives the compiler some extra guarantees that it can use to optimize things behind the scenes. Even if the compiler just ends up creating a future behind the scenes, that implementation is completely hidden to the programmer, and can be updated/changed at any time by the compiler team without involving evolution.

Let’s say at some point down the road as part of the actor work, the compiler ends up getting some visibility into how actor queues will be coalesced… Then it can use that knowledge to reorganize the code above to be more efficient.  For example, if it knows that the part of the calls to fetchImage() will be serialized by an actor, it can just make that part synchronous and avoid the asynchronous overhead.  If we had baked in everything at the low-level to use future objects, we wouldn’t be able to make that optimization.

There are 2 big results from this:

1) It gives us the ability to easily start several things behind the scenes and then await the group of them finishing

2) Our implementation of Futures (and other constructs) can potentially take advantage of the compiler optimizations by using this beneath the hood.

The first point is huge for me. This is an extremely common thing to want in real-world code, but it usually requires a bunch of complex (sometimes error-prone) machinery to actually pull off. There is often wasted overhead as well.  I like that I can do it here in a very natural and lightweight way.  (Futures also simplify this problem a great deal, but I wouldn’t really call them lightweight). 

> You keep bringing up this suggestion, so I must be missing something, but it seems to me that your suggestion is covered by Futures. Why is calling with 'async' better?

As I said above, I think this will help our Futures be much more efficient when we build them, but there are also some advantages of having this as an option and not just using Futures for everything.

Mainly, futures require a different return type, and they require boxing/unboxing.  To chain futures, each link in the chain has to be future-aware… and each link can only be used with futures (i.e. the return type is Future<MyType>).  There are a lot of cases where I am boxing just to unbox and re-box.  Most of that is fine as long as the framework was built with it in mind, but it doesn’t work as well if I want to add the functionality post-hoc.

With ‘async’ being used to defer ‘await’, it just works on any function which can be awaited.  Further, it can be chained naturally through any number of async functions.  I showed this in another thread, but if you think through what happens when you return something which still needs to be awaited, you see that that state is naturally passed along with it (the function was either called with ‘await’ or ‘async’, and thus the return value is either awaited on, or the result picks up the compiler marking from ‘async’ showing it still needs to be awaited).  Because there is no (visible) boxing, you don’t have to worry about monads, etc…

It basically nests in the same way that try/throws nests, and we already have to do that for async/await anyway… so it is a lot of win for very little additional complexity (at least on the usage side).


> -BJ
> On Sep 3, 2017, at 6:01 PM, Jonathan Hull via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>> On Sep 3, 2017, at 9:04 AM, Chris Lattner via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>> On Sep 3, 2017, at 4:00 AM, David Hart <david at hartbit.com <mailto:david at hartbit.com>> wrote:
>>>>> Please don’t read too much into the beginAsync API.  It is merely a strawman, and intended to be a low-level API that higher level abstractions (like a decent futures API) can be built on top of.  I think it is important to have some sort of primitive low-level API that is independent of higher level abstractions like Futures.
>>>>> This is all a way of saying “yes, having something like you propose makes sense” but that it should be part of the Futures API, which is outside the scope of the async/await proposal.
>>>> But it would be nice for all high-level APIs that conform to a Awaitable protocol to be used with await without having to reach for a get property or something similar everytime.
>>> The futures API that is outlined in the proposal is just an example, it isn’t a concrete pitch for a specific API.  There are a bunch of improvements that can (and should) be made to it, it is just that a futures API should be the subject of a follow-on proposal to the basic async/await mechanics.
>> Would it be possible to have the manifesto be a series of proposals then?  I really think it is important for us to look at how all of these things fit together.  I agree that async/await should come first, but looking at how concrete things like Futures would work may help to inform the design of async/await.  We should do the back-propigation in our design before anything is locked in...
>> The thing I would most like to see as a quick follow-on to async/await is the ability to use the ‘async’ keyword to defer ‘await’. This feels very natural, is highly optimizable by the compiler, and it allows for a lot of very common use-cases which are not covered well by pure async/await.  I think it would have a large impact on the eventual design/implementation of futures (and at least some impact on the design of async/await).
>> Thanks,
>> Jon
>> _______________________________________________
>> swift-evolution mailing list
>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170903/b796e381/attachment.html>

More information about the swift-evolution mailing list