This article isn't going to teach you about the latest trends in frontend development. Or look in detail into the way to get the most performance o...
For further actions, you may consider blocking this person and/or reporting abuse
IMO, the part about svelte's reactive language falling short is only half-truth. In svelte if we want composition we'd start with writable store in the first place, with practically the same reactive language like Vue's.
Giving the
$store
syntax sugar, it feels very native to svelte's reactive language, and shouldn't be left unmentioned.Talking about store being auxiliary, it's actually a good thing, not some kind of burden/cost. Cus it's opt-in, swappable. You can freely switch to redux or rxjs store if you want. You don't get eco locked-in like with Vue's
$ref
or Solid'screateSignal
.Right, but as mentioned a couple of times, that is outside of the "language" part. I like Svelte stores. And they solve a very very necessary problem and having sugar makes them feel more native. But the juxtaposition makes it instantly clear of language/compiler limitations. It wraps a 2nd completely different reactive system. If Svelte only used stores I suspect it might not have been so. The insistence on being just JS/HTML by its followers also amplify this.
And really the purpose of this article is wondering if we can somehow find the holy grail. A truly composable reactive system that doesn't introduce a ton of new syntax. Svelte gets most of the way there, but what does all the way look like?
Not sure how holy anything I make can be, but my reactive system hides pretty well behind "define_property": tilton.medium.com/simplejx-aweb-un...
I have been enjoying your surveys of reactive alternatives and learned a few new ones! I need to get out more, missed Recoil completely!
The problem may be in trying to make standalone vars reactive. But standalone vars may be a requirement because objects got thrown out, because a reactive object can hide reactivity behind reactive accessors. So the real problem, then, is throwing out objects. The funny thing being that objects where different instances can have different reactive definitions for the same property, and where different instances can have additional properties (the prototype model), kinda solves all the problems of OO. But we cannot have objects because ReactJS (how ironic now is that name?) needs functions everywhere so they can control state change, which is also why React add-ons must now let React manage all state. See "concurrent mode". So the real problem may be the impedance mismatch between trying to achieve reactive state in ReactJS, which has officially rejected the paradigm for its eagerness.
the tl;dr for the above is "slippery slope". :)
Idea about double-destiny operator:
This would be impossible to implement in the general case because the compiler would essentially be solving an equation. In most cases, this would be impossible because there could be multiple solutions for a depending on the value of b. In other cases, reversing the operation would simply be impossible, e.g. a hash function.
How about a new keyword?
Then we are looking at a label/keyword example and that all applies. My point is it is easy to bucket all solutions into these categories or hybrids of them. We can debate the exact naming/syntax but I have been wondering if we can escape this altogether.
That's interesting. Scares me a bit. Part of me really wants to make mutation(assignment) special and isolated but I just might not be letting go enough to fully embrace this thinking. Would it ever be difficult to reverse the derivations? Sure subtracting 1 from b is easy enough. I just wonder if that wouldn't always be the case.
It's the lens in general. See JS example:
We actively use this that approach in this way:
It only makes sense if the mapping is a bijection (math term). It's a really rare property, meaning zero information loss.
No, It's very common. We can bind local property with part of json which bind with local storage as example. So write to property will change json at local storage and affects to same property of another instance of same app. Example:
Bi-directional bindings re-invented? Fair enough.
FWIW, the Vue example is a bit misleading: you don't need
$$()
for common reactive effects if usingwatchEffect
:$$()
is only needed if you have external composition functions that explicitly expect a raw ref object as arguments.Ok. Thanks Evan. I have updated that section to better represent common patterns in Vue. Thank you.
For what it's worth, without completely messing with the semantics I think the Function decoration approach like found in Vue Ref Sugar is the only approach that actually checks all the boxes. But I'm interested in what happens if we do just mess with everything.
These are all interesting experimentations. And here's one approach I've been explororing since last year: Subscript - reactivity without any special syntaxes or language tokens, but just regular, valid JavaScript.
It is implemented as a UI binding language:
Above, the 'this' keyword is a reference to the
<div>
element; andthis.state.title
is the state being observed. Now thelet
expression evaluates each time the state ofthis.state.title
changes, and thethis.setAttribute()
call picks up the new value oftitle
each time. This is what happens when state is changed as in below:It's that simple; it's pure JavaScript that works reactively by just understanding reference stacks. Details are here: webqit.io/tooling/oohtml/docs/gett...
I moved on implementing it's runtime and making a real world app with it. Waiting to see where this leads.
I see in this example you have accessor on the state on the element which serves as the signal and the script("subscript") itself is the wrapped effect. Makes sense. I see no problem with runtime reactivity without special syntax. SolidJS works that way today. But the desire for getting rid of the
this.___
orstate.___
is almost feverish pitch so I thought I'd try my hand at the problem.It is interesting to see here the destiny operator, in which we have been using component composition description language for a long time.
A few examples of how we declare object methods (channels):
This is a method that returns a constant string (one way channel).
This is the same, but the meaning can be changed (singal in your terminology and two-way channel in ours).
Now the operator of destiny:
There is already a derived method that returns an array from a constant string and the values of another method.
And now, the most interesting thing is the bidirectional channel:
We can read and write in 'title' without even knowing that we are actually working with 'name'.
And then there is the reverse destiny operator:
It may seem that this is the same as the normal destiny operator, but it matters when linking components to each other:
Now the
Output
directly uses the value from theInput
, and we control it through the "name" channel.And yes,
Input
andOutput
is channels too which returns cached instance of other components.Here you can see the generated TS code.
I love the concepts. For these new concepts to be adoptable into EcmaScript they'd have to play well with the existing imperative constructs, living along aside them, while reducing any chance for ambiguity.
Maybe the label idea isn't so bad if it propagates into every place it the feature is used, like
or something.
Now, I'm not sure this is the best syntax, or that it doesn't have any issues, or that
@
is the best symbol, but the idea with signal names requiring to be postfixed with@
in the example isAnother thing to consider is that, if dependency-tracking were applied to all of JavaScript's existing features, what would this imply for performance?
The performance characteristic of every variable, every property, every parameter, every usage site, would change (probably get slower) just so that reactivity works. With a parallel syntax to keep the imperative and declarative paradigms decoupled, we can limit any overhead strictly to those signal variables and their consumers, without affecting the engine implementation of the other features. This would reduce implementation complexity for people who write and maintain the JS engines.
I'm hoping this will spawn more discussion on the possibilities!
I've played around this topic a bit myself, and the gist was something like this:
This basically lets one compose values as effects too, example:
My thinking is that a variable that effects from others won't ever directly change itself, so that setting
c(value)
might, instead, throw an error.As for the syntax, I find the reactive bit being well represented by functions so that
let b <- a + 3;
doesn't look too bad:() => a + 3;
except it accepts zero arguments as it cannot be directly invoked, and the arrow points at the reference that should reflect whatever the body/block returns.Interesting stuff. I wrote a language that I never got to implement (of course but who knows if one day I won't find the time to do it) that does what you suggest. The language is simple, and based on the core functional reactive programming concepts. You have events, effects, and pure computations. Excerpts:
So just three key elements of syntax,
event => action
notation to express reactions,=
for equations that must always hold at any point of time (yourconst
?), and<-
to update an equational variable.There is nothing particularly smart there. Synchronous reactive languages are not much different than that (Lucid, Esterel, etc.). It is not the smartness, it is more the simplicity of the notation.
For composition, it is simply about noting that a program is a set of events (left side of x => y), variables (left side of x = y), and action/effect handlers (right side of x => y). So
Program<Events, Variables, Effects>
basically. To get one big Program from two small Programs, just import those two small in the big one and possibly rename the variables, events, effects (to avoid shadowing) - just like you would do with template partials in good old HTML templating systems. But the story was not perfect for composition. Renaming is a hassle that breaks a bit module independence (the big component that reuses the small one need to know about the details of the used component. Ideally you should not need to know about the variables in the reused component, they should be encapsulated).Haven't put myself to solve these problems though. So it is interesting to read this writing.
When we talk about reactive language we should think not about reactive programming or even reactive data accessors, but about reactive data structures. The main point of separate language for reactive datas is that it should work perfectly (effecianlty) with a data with all possible (by a language) ways. In other words, the language should expose API for data accesing and transformation and limit it to fit in a most efficient compile output. It means we should be able to analize AOT all accessors in classic
for
/map
and optimize it or throw it away from a language and replace it by some variations ofpick
:listOfAC = listOfAB.pick({a: 'identity', b: b => toC(b) })
.I think there probably is some amount of API that is unavoidable but there is something attractive about not having much in the way of API for this. We might have no choice for things like lists. To specially handle compilations for things like array literals. Mostly I don't view this necessarily always been in a runtime primitive mechanism. The reason I focus on language is because as our analysis gets better the implementation could change dramatically. We see this with Svelte and I've seen this taken even further with Marko. Those don't even have reactive primitives anymore at runtime but just call some functions. My hope is that behavior (but not the implementation) can be well defined in mostly with regular language mechanisms. Lists might just have to be the exception.
Svelte
?:
Solid
!: