-
Notifications
You must be signed in to change notification settings - Fork 68
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Update returns Command #282
base: master
Are you sure you want to change the base?
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I've only read the core part of this, but have enough questions to start a conversation :)
pub fn render<Event>(&self) -> Command<Event> { | ||
Command::empty_effect(self.render_async()) | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do I imagine this correctly that the constraint which ensures that effects coming out of an update
function only ever result in Events the update function is able to process (potentially wrapped with a parent app Event) has moved from being an argument of the capability itself to being tied to the return value of the capability's API?
That hopefully means the bound is still strict and we don't allow App
s to create effects resulting in events of any type, except their own.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I assume it also means that Commands
can be "lifted" to a parent app (meaning their events will be auto-wrapped with a given event)?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
(future me: yes, they can)
pub fn render_async(&self) -> impl Future<Output = ()> { | ||
let ctx = self.context.clone(); | ||
self.context.spawn(async move { | ||
async move { | ||
ctx.notify_shell(RenderOperation).await; | ||
}); | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I wonder whether there's a way to avoid this shell side-channel altogether.
Is there a way we could have a specific future type for each cardinality of the shell interaction (notify/request/subscribe), which exposes the operation it was created with, so that the operation accessible to the executor?
I can't quite see it clearly enough, but I think it would allow the executor to be aware of these "special futures" and whenever the tasks yield because they are awaiting on these special futures, the executor would know how to
a) get the operation out and send it to the shell
b) resolve the operation on that future so that it can proceed, which should result on them polling as ready
That should theoretically eliminate this channel altogether. The futures would just be a kind of "yield point carrying the value that needs processing". (That might make them more easily visible to tests as well)
I'm not totally sure this is possible to implement, but if it is, it would mean capabilities would be fully free of context.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thinking about it a bit more - the problem with that idea is that those special futures will be inside one of the states of an anonymous future made by an async
block, which is opaque, isn't it... the "innermost" future needs some way of exposing the operation that can resolve it to the outside.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
But prehaps this could be done on a more local scale within the Command
. The command could take a function which gets an equivalent of the context provided by the executor later. That context would just have a local channel back up to the command, allowing to bypass the anonymous future layers.... or something along those lines.
pub(crate) struct QueuingExecutor<Event> { | ||
ready_queue: Receiver<TaskId>, | ||
ready_sender: Sender<TaskId>, | ||
tasks: Mutex<Slab<Option<BoxFuture>>>, | ||
tasks: Mutex<Slab<Option<BoxFuture<Event>>>>, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The Event
ends up being the top-level app's Event
type, presumably?
Missing, | ||
Unavailable, | ||
Suspended, | ||
Completed, | ||
Completed(Command<Event>), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm not sure I follow this part - why does completed task hold aCommand
? Is it the command that started it? or the command that is to be processed by the shell...? I'm not totally sure what's going on 😅
I would've thought the task is either just done, or if anything it's done and (optionally) resulted in an Event
...?
I suppose the granularity of tasks may now be different than before, or maybe I'm just lost 🤔
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
When a task resolves it always returns a command.
Think of the existing model - e.g. an HTTP request that resolves to an event. When the request resolves, the resulting Event is implicitly spawned onto the executor through the spawn_event
channel that is part of the capability context.
With the changes in this PR, the spawn_event
channel has gone, and the event (or any other kind of Command
) is returned directly from the resolved future. The command needs to be explicitly handled - either fed into app.update
if it is an event, or spawned onto the executor if it is an effect (or else the command may be Command::none()
in which case there is nothing left to do).
This happens in code via an interplay between Core::process
and QueueingExecutor::run_all
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Wait, are you saying the task completes every time the code in the future(s) hits an .await
on one of the Crux specific interactions?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
No, only when it resolves. i.e. the future is polled to completion (potentially through multuple .await
s). Uh, I suppose in crux
we are using 'resolve' to mean something else (Update::resolve
means to resolve an async effect) so maybe not a good word to use.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So then I'm back to the question of how does it complete with a Command
which is not just an Event
?
I'm not sure what that would mean in principle. The parent future of the effect task has finished, there are no more effects to do, the only possible outcomes are either nothing (we're done) or an event back into the app, right? What am I missing?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It could be made to return Option<Event>
instead, but you would lose some expressivity.
A Command
allows one to return any combination of events and effects to run next. This is useful e.g. to do an http request and a render call concurrently. You could do them in a single future with a future::join
call, but then they are coupled together. Returning both with Command::effect(render_fut).join(http_fut)
decouples them.
example:
crux/examples/cat_facts/shared/src/app.rs
Lines 143 to 156 in 03a7c64
let eff1 = caps | |
.http | |
.get(FACT_API_URL) | |
.expect_json() | |
.send_and_respond(Event::SetFact); | |
let eff2 = caps | |
.http | |
.get(IMAGE_API_URL) | |
.expect_json() | |
.send_and_respond(Event::SetImage); | |
let eff3 = caps.render.render(); | |
eff1.join(eff2).join(eff3) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actually possibly talking cross-purposes here. The above applies to the return type for fn update(..)
. You want to know why any spawned future must return Command
? The answer is it seems reasonable that you might want to spawn a future (or several futures) from within another future (remember we have no way to spawn futures other than to return them, as context.spawn
has gone).
You might argue that futures shouldn't attempt to spawn futures directly, they should return events which subsequently kick off the futures. Maybe true? Returning a Command
seems to give maximum flexibility
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I seee - if you wanted to start a separate task that runs concurrently without awaiting the result. Yea, that is something we need o retain as an ability...
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I suggest staring at
crux/examples/counter/shared/src/capabilities/sse.rs
Lines 34 to 65 in 03a7c64
pub fn get_json<F, Ev, T>(&self, url: impl AsRef<str>, make_event: F) -> Command<Ev> | |
where | |
F: Fn(T) -> Ev + Clone + Send + 'static, | |
Ev: Send + 'static, | |
T: DeserializeOwned, | |
{ | |
let context = self.context.clone(); | |
let url = url.as_ref().to_string(); | |
let stream = context | |
.stream_from_shell(SseRequest { url }) | |
.flat_map(move |response| match response { | |
SseResponse::Chunk(data) => { | |
// this decoder is async (even though for our purposed it doesn't need to be) | |
// which makes the following code a bit fiddly | |
let reader = decode(Cursor::new(data)); | |
let make_event = make_event.clone(); | |
let inner = reader.map(move |sse_event| { | |
if let Ok(Event::Message(msg)) = sse_event { | |
let t: T = serde_json::from_slice(msg.data()).unwrap(); | |
let make_event = make_event.clone(); | |
Command::event(make_event(t)) | |
} else { | |
Command::none() | |
} | |
}); | |
Box::pin(inner) as futures::stream::BoxStream<_> | |
} | |
SseResponse::Done => Box::pin(futures::stream::empty()), | |
}); | |
Command::stream(stream) | |
} |
It's a stream where each output is itself a stream of Command
s
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It's a stream where each output is itself a stream of Commands
That seems a step too far 😅
I think the problem there starts with insisting on Event
being a return, or generally the command async code working towards a final return value. I don't know that's the most straightforward approach to it.
The async code in capabilities (and now in Command
) is quite intentionally pretty imperative, because it's "process" code, rather than a computation (as in - more Turing, less Church). It's much more naturally a "do this, then do that, then do that, then finish". And yes we can model that with a recursion, but as can be seen in the code above^ it's really not very straightforward that way. We could do better with some monadic style combinators like Command::event(event).and_then(...)
but... I think async is just a nicer interface for this type of work.
|
||
let future = { | ||
let counter = counter.clone(); | ||
async move { | ||
assert_eq!(Arc::strong_count(&counter), 2); | ||
ShellRequest::<()>::new().await; | ||
Command::none() |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
🤔🤔
crux_core/src/lib.rs
Outdated
pub fn map<F, Event2>(self, func: F) -> Command<Event2> | ||
where | ||
F: Fn(Event) -> Event2 + Send + Clone + 'static, | ||
Event: 'static, | ||
{ | ||
let inner = match self.inner { | ||
CommandInner::None => CommandInner::None, | ||
CommandInner::Event(ev) => CommandInner::Event(func(ev)), | ||
CommandInner::Effects(effects) => { | ||
let effects = effects | ||
.into_iter() | ||
.map(|fut| { | ||
Box::pin(fut.map({ | ||
let f = func.clone(); | ||
move |cmd| cmd.map(f) | ||
})) as BoxFuture<'static, Command<Event2>> | ||
}) | ||
.collect(); | ||
CommandInner::Effects(effects) | ||
} | ||
}; | ||
Command { inner } | ||
} | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Noice.
examples/cat_facts/shared/src/app.rs
Outdated
let _result = caps.key_value.set_async(KEY.to_string(), bytes).await; | ||
Command::none() | ||
}) | ||
.join(async move { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
presumably this could just be
.join(caps.render.render())
or is this just to show that either is valid?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I have reworked the Commands and effects to make this cleaner
I have merged in |
I have talked before that I don't like how
fn update(..)
does not return anything, all the effects are spawned implicitly by the capabilities. I thought I'd see whether I could make it more Elm-like and return aCommand
instead.Turns out to be quite invasive but I do think it is an improvement. Especially, taking the
Ev
generic out of every Capability is nice. And totally removing the need for aCompose
capability.