Discussion on: What happened to Components being just a visual thing?

[peerreynders]

I'm only concerned with the application, the architecture , if possible, in a ideal world.

But how are you going to justify the view/domain separation to a component die-hard? They are going to ask "what's in it for me?"

"The web" is too specific.

"The web" also has a habit of scuttling good ideas with its unique challenges (unless they were carefully considered up front).

---

In a way you have landed on the Humble Dialog (or Humble Object) from 20 years ago.

I agree with your analysis that it makes no sense to organize the "application" in terms of the flow of elements in the document tree; I think the heavy use of Context and external state managers in React applications is testament to that.

That being said I think that it may be necessary to avoid the term/concept Component altogether when describing a new architecture because the intended audience will most likely use React Component as a reference and assume its particular compositional qualities (and capabilities with regards to view/application state).

Domain logic partitions around domain capabilities while the UI partitions around UX concerns - both of which may or may not align.

---

Another aspect is that developers of component-centric designs may just favour bottom-up design

Quote:
"In order to get IntelliSense to work correctly, bottom-up programming is best. IntelliSense wants every class, every method, every property, every field, every method parameter, every local variable properly defined before you refer to it. If that’s not the case, then IntelliSense will try to correct what you’re typing by using something that has been defined, and which is probably just plain wrong."

By being concerned about "application architecture" (top-down) you're in the app-centric camp—while meaningful boundaries still exist, component boundaries aren't the salient concern.

---

I assume you are familiar with Martin Fowler's GUI Architectures Page and André Staltz's Unidirectional User Interface Architectures.

[3Shain]

But how are you going to justify the view/domain separation to a component die-hard? They are going to ask "what's in it for me?"

TBH I haven't thought about this. I thought it's natural to choose the former because it's better architected (hope it's not a bias but can be theoretically proved). Although It's never shown that people will always choose the "better solution in theory" ,but likely the "good solution that fits my problem".

"The web" also has a habit of scuttling good ideas with its unique challenges (unless they were carefully considered up front).

(handclaps)

---

Another aspect is that developers of component-centric designs may just favour bottom-up design
By being concerned about "application architecture" (top-down) you're in the app-centric camp—while meaningful boundaries still exist, component boundaries aren't the salient concern.

Recently I realised I was finding (or has found) another "Component" abstraction which is different from the current Visual Component (as I really want to preserve the fractal structure). But yeah it might be better to give it another name other than Components.

---

I did read about the materials you've mentioned. In fact I knew about them from your previous comments and I have to say I've learned a lot. Thank you for your always informative sharing. I hope one day I can show an opinion always with dense external sources like you.

[peerreynders]

Although It's never shown that people will always choose the "better solution in theory", but likely the "good solution that fits my problem".

When was the last time you were presented with a counter app like this?

import { createRoot } from 'react-dom';
import { createContext, useContext, useEffect, useState } from 'react';

function makeCountApp(count) {
  const listeners = new Set();

  return {
    increment,
    subscribe,
  };

  function notify(listener) {
    listener(count);
  }

  function notifyListeners() {
    listeners.forEach(notify);
  }

  function increment() {
    count += 1;
    notifyListeners();
  }

  function subscribe(newListener, now = false) {
    listeners.add(newListener);
    if (now) newListener(count);
    return () => {
      listeners.delete(newListener);
    };
  }
}

const AppContext = createContext();

function UI({ initCount }) {
  const [count, setCount] = useState(initCount);
  const app = useContext(AppContext);
  useEffect(() => {
    const update = (newCount) => setCount(newCount);
    return app.subscribe(update);
  }, [app]);

  return (
    <div>
      <p>You clicked {count} times</p>
      <button onClick={app.increment}>Click me</button>
    </div>
  );
}

function App() {
  const initCount = 0;
  return (
    <AppContext.Provider value={makeCountApp(initCount)}>
      <UI initCount={initCount} />
    </AppContext.Provider>
  );
}

const rootElement = document.getElementById('root');
const root = createRoot(rootElement);
root.render(<App />);

which I would classify as app-centric; versus this (component-centric):

import { createRoot } from 'react-dom';
import { useCallback, useState } from 'react';

const increment = (value) => value + 1;

function App() {
  const [count, setCount] = useState(0);
  const incrementCount = useCallback(() => {
    setCount(increment);
  }, []);

  return (
    <div>
      <p>You clicked {count} times</p>
      <button onClick={incrementCount}>Click me</button>
    </div>
  );
}

const rootElement = document.getElementById('root');
const root = createRoot(rootElement);
root.render(<App />)

Many people gravitate towards what's easier and expedient in the short term ("the scam") and that is what the "business" often wants—faster time to market. Your don't have to go to apply for loan to go into technical debt, you just do it. Once React was described as the "V in MVC" but from I can tell components (with hooks) are often kitchen sinks combining various UI and application responsibilities-the only mitigating factor being that the complexity is limited to a single UI Element or the complexity is delegated to nested components (reminding me of HMVC). And the culture in the community seems to favour component-centric approaches (e.g. Application State Management with React).

Perhaps if you are going to rewrite the product every two years anyway it doesn't matter that there is no discernible "architecture", however it may also make the rewrite a foregone conclusion.

"Software architecture is those decisions which are both important and hard to change" (Making Architecture Matter). The flip-side is that failing to identify up front what is truly important and what will need to change in the future can still lead to poor architectural decisions.

I did read about the materials you've mentioned.

Native app developers seem to focus on MVVM, MVP(VM) or VIPER these days. 43 years have passed since the introduction of MVC and the refined UI patterns still keep coming. This illustrates that this is a non-trivial problem without a one-size-fits-all solution. I think it's important to be familiar with the various ways this problem can be broken down—that way simple problems can be solved simply, while options (which are overkill otherwise) exist for the more gnarly cases.

---

FYI: ThoughtWorks Technology Radar: SPA by default:

"Too often, though, we don't see teams making that trade-off analysis, blindly accepting the complexity of SPAs by default even when the business needs don't justify it. Indeed, we've started to notice that many newer developers aren't even aware of an alternative approach, as they've spent their entire career in a framework like React."

[3Shain]

---

When was the last time you were presented with a counter app like this?

I was crafting a library/framework/architecture(still in private yet) and this is how it works:

initially it looks like typical "component-centric"

import { State, __buildAppNotStableYet } from 'kairo';
import { UI, Component } from '@kairo/react';
import React from 'react';
import { createRoot } from 'react-dom';

const increment = (value) => value + 1;

const Counter = UI(function*() {
    const [count, setCount] = yield* State(0);
    const incrementCount = ()=>setCount(increment) 
    return Component((_,$) =>
    <div>
      <p>You clicked {$(count)} times</p>
      <button onClick={incrementCount}>Click me</button>
    </div>);
});
// typeof<Counter> : ImplementationOf<'@kairo/react:UI',DependsOn<'CounterApp'> | WithLifecycle, {}, FunctionComponent<{}>>

const { instance: App, start } = __buildAppNotStableYet(
    Counter.provide(AppImpl) // assembly your application
);
// typeof<App> : FunctionComponent<{}>
start();
const rootElement = document.getElementById('root');
const root = createRoot(rootElement);
root.render(<App />);

but soon you could refactor it to a more decoupled version (you could say it's "application-centric" coz you can write App and unit test it regardless of UI, the same is true for UI)

import { State, createConcern, Cell, __buildAppNotStableYet } from 'kairo';
import { UI, Component } from '@kairo/react';
import React from 'react';
import { createRoot } from 'react-dom';

const increment = (value) => value + 1;

const App = createConcern<{
    count: Cell<number>,
    increment: ()=>void
}>()('CounterApp'); // a branded interface, also a injection token

const AppImpl = App(function*({initCount}:{initCount:number}) {
    const [count, setCount] = yield* State(initCount);
    const incrementCount = ()=>setCount(increment);
    return {
        count,
        increment: incrementCount
    }
}); // implement the interface (concern)
// typeof<AppImpl>: ImplementationOf<'CounterApp',WithLifecycle,{ initCount: number }, {
//  count: Cell<number>,
//  increment: ()=>void
//>

const Counter = UI(function*() {
    const { count, increment } = yield* App;
    return Component((_,$) =>
    <div>
      <p>You clicked {$(count)} times</p>
      <button onClick={increment}>Click me</button>
    </div>);
});
// typeof<Counter> ImplementationOf<'@kairo/react:UI',DependsOn<'CounterApp'>,{}, FunctionComponent<{}>>

const { instance: AppComponent, start } = __buildAppNotStableYet(
    Counter.provide(AppImpl.withProps({initCount: 0})) // assemble your application
);
// typeof<AppComponent> : FunctionComponent<{}>
start();
const rootElement = document.getElementById('root');
const root = createRoot(rootElement);
root.render(<AppComponent />);

That's my solution anyway. How is it different from current Component model? I'm arguing current Components are too bloated so everything kairo provided are primitives - do one thing only and it encourage you to decompose things into small concerns (if applicable). For UI (frameworks like react) the (visual) Component should be only visual thing.
One could argue "then everything is coupled with kairo" then I dare say I'm shipping language-level abstractions/primitives (that might be biased).
(and an ultimate goal is to supporting all mainstream UI frameworks(as well as some leaner render coz a full Component is not required), and they are supposed to truly play the role of view layer. maybe even not limited to front-end)

Don't blame that it doesn't focus on web, it's never supposed to be. SPA be abusing doesn't matter that (Visual) Component-based is a sub-optimal architecture.

Side note: generator is used to provide DI mechanism and static type checking (that the type system can tell you what dependencies are missing). Notably I didn't use react context. For hierarchy structure I mentioned this a little bit in my previous comments.

[peerreynders]

Interesting approach; I have to admit I didn't even know yield* was a thing.

that the type system can tell you what dependencies are missing

Yes but given IteratorResult<T, TReturn> my impression is that T is going to be union of the dependency types-so how would TypeScript check for the correct cardinality and ordinality of the dependencies? I also suspect that a minority of Developers are fluent with generators so that may present an additional cognitive barrier.

My personal experience has been that generators can be slow so it would be unfortunate if that would impose a premature performance ceiling; apart from the concern that generators are a runtime mechanism that may prove impossible to streamline with some design/compile time generated injection scripts.

But I'm sure I'm missing something.

One could argue "then everything is coupled with kairo"

• Has it dependencies on a particular JavaScript runtime (e.g. browser vs. node; hopefully not)?
• Does it have the potential to slow down microtests (> 100ms)?
• Does it impede the ability to test relative small units of capability (hopefully not)?
• It really boils down to how easy (and fast) it is to test (I'm not keen on the React/Jest situation).

Would const [count, setCount] = yield* State(initCount); actually work if it was run/tested independent from React? (seems UI would have to thread useState into App; one of my beefs with hooks it that they rely on a React specific mechanism).

Those are just my first impressions.

[3Shain]

Yes but given IteratorResult my impression is that T is going to be union of the dependency types-so how would TypeScript check for the correct cardinality and ordinality of the dependencies?

Thant's why I'm using branded interface (nominal typing) so that I can 'add' or 'subtract' from a union of literal string (or object with literal string type field).

I also suspect that a minority of Developers are fluent with generators so that may present an additional cognitive barrier.

I thought it's not typical generator function but more like a DSL. And in theory it's closer to Algebraic Effects (or Effect Handling) so that generator only provides the capability to interrupt the control flow. Just borrowing the syntax, I'm not expecting (dev) users to fully understand generator (unless they want to hack into low level implementations).

My personal experience has been that generators can be slow so it would be unfortunate if that would impose a premature performance ceiling

Only one-shot execution. Much better than react's repetitive render call (and repetitive closure creation).

apart from the concern that generators are a runtime mechanism that may prove impossible to streamline with some design/compile time generated injection scripts.

they are likely different concerns. i'm not requiring devs to replace their every function with generator ones.

Has it dependencies on a particular JavaScript runtime (e.g. browser vs. node; hopefully not)?

purely ecmascript standard, purely runtime mechanisms.

Does it have the potential to slow down microtests (> 100ms)?

negative. if generator is thought to slow down executions, they only affect the construction process (which only happen once). I don't believe this will dramatically affect the performance.

Does it impede the ability to test relative small units of capability (hopefully not)?

not at all. it's one of the design goal.

It really boils down to how easy (and fast) it is to test (I'm not keen on the React/Jest situation).

it should be as easy as test against a regular class object (after all I'm utilising generator to implement an atypical "constructor" that handles DI and provide dependency type inference).

Would const [count, setCount] = yield* State(initCount); actually work if it was run/tested independent from React? (seems UI would have to thread useState into App; one of my beefs with hooks it that they rely on a React specific mechanism).

exactly. it's a stand-alone fine-grained reactivity impl (similar to solidjs)

it's the Component((_,$)=>...) API connecting two worlds. UI is a branded interface for React.FunctionComponent (from a top-down view, UI is the last (ultimate) concern (matches UI as an afterthought); from a down-top view, UI is the most concrete thing you can start with, and you can implement everything inside initially then move things out gradually). React is agnostic to anything else.

---

I'm think about a proper naming of this kind of framework because it's different from any kinds of existing ones. I had a idea kernel framework (from Functional Core Imperative Shell, that the UI frameworks frame the shell while kairo frames the core)

[Magne]

I'm not very sure it's the same idea I've talk about, I think it is. The idea is prevent business logic from being leaked into the view.

Yes, I think it’s precisely that same idea. Pete Heard talks a lot about separating out all business logic, so it is testable, and the UI being an afterthought (even one step further than what MobX does).

[peerreynders]

In broad strokes he's mapping Clean Architecture to the front end.

I'm still partial to this 2018 "actor inspired" architecture for the front end.

• Each "actor" is a message processor which maintains it's own autonomous state.

• It doesn't share anything and only runs when it has a message to process

  • It processes any one message as quickly as possible while sending messages to other actors (possibly spawning new ones).

• “An Actor, a model and an architect walk onto the web...”

• The 9am rush hour

• Lights, Camera, Action!

• Headless Web Development

• When should you be using Web Workers?

PolymerLabs / actor-boilerplate

A starting point for web apps based on the actor model.

Actor Boilerplate

A starting point for web apps based on the actor model.

Chrome Dev Summit 2018 Talk

Architecting Web Apps - Lights, Camera, Action!

We also wrote a series of blog posts with more detail on web development with the actor model:

• The 9am Rush Hour
• Lights, Camera, Action!
• Headless Web Development

What is this repository that am I looking at?

This is a very basic web app that uses the actor model. The actor model helps you to break your app’s core logic into small pieces that have to communicate with messages instead of using function calls. Adopting this model has multiple benefits on the web:

• Yields to browser (it naturally leads to chunked code)
• Encourages lazy-loading and code splitting
• Gives you a clear separation of concerns
• Makes moving code off-main-thread easier
• Resilience against unexpected long-running tasks
• Enables multi-modality for web apps

What’s in here?

This boilerplate…

View on GitHub

PolymerLabs / actor-helpers

actor-helpers

Helpers to build web applications based on the actor model. These helpers are used in our examples, for which you can find our boilerplate here. We encourage you to read through the boilerplate examples first and then read through the code in this repository.

actor.ts

actor.ts contains a base class implementation for an actor, as well as functions to hookup() and lookup() actors on the web page. For more detailed examples, please check out the in-file documentation.

watchable-message-store.ts

This store is an implementation detail of the messaging system used by hookup() and lookup() to allow actors to communicate with one another. You shouldn't need to interact directly with the message store, but it's here all the same if you do.

---

License BSD-3-clause

Please note: this is not a Google product.

View on GitHub

Too bad it didn't go anywhere … but as usual I digress …

[Magne]

What do you think about XState, which is actor model inspired state handling often used on the front-end?

Ideally, I imagine it would be nice with Niladic Components (no props), so React could be a dumb View layer, and simply handle all state separately in XState…

[peerreynders]

See my comment here

Expressed roughly in JavaScript-like syntax an Erlang process is at its simplest

function loop(state) {
  const msgFrom = receive();

  const nextState = (lastState, { type }) => {
    switch type {
      case 'a':
        send(msgTypeA);
        return toStateA(lastState)

      case 'b':
        send(msgTypeB);
        return toStateB(lastState)

     default:
        return lastState
    }
  }(state, msg);

  loop(nextState)
}

Of course JavaScript

• doesn't implement tail(/last) call optimization
• isn't pre-emptively scheduled. Typically a process is blocked at a receive() (or any other blocking call) waiting for the next message and even if it isn't the scheduler will suspend it after it consumes it's reductions.

None of this is feasible in single threaded/event loop based JavaScript but it points to the possibility of tying process scheduling to message delivery:

• A "processing scope" (replacing "actor") only gets processing control when it receives a message (with some exceptions; the message would likely be arrive via a subscription). To be well behaved it has to break up large tasks into phases which it initiates by sending a message to itself at the end of the current phase.
• To interact with other "processing scopes" is has to send a message. However there is no immediate response (the message goes to the end of the delivery queue). If there is a response it will be received later at which point in time it can be processed (so the processing scope state has to reflect this "in progress interaction" - the equivalent to a "function call (returning a value)" to a separate processing scope).

So the basic functionality needed for a processing scope is a way to receive messages so it has something to do, and a way to send messages so it can collaborate with other processing scopes. For that to work the message router/scheduler needs to support a few things:

• The entire system is built around postMessage(). For the purpose of discussion lets call a Window or Worker a "processing node".
• Each node
  • routes messages (i.e. has to manage routing data)
    • messages to the local processing scopes are put on the delivery queue
    • messages to processing scopes managed by another node are immediately posted to the nearest node.
  • schedules processing by
    • delivering a single message to the local processing scope from the head of the delivery queue provided there is enough time left in the current processing slice
    • returning control to the event loop when the current processing slice is exhausted.
      • if the delivery queue isn't empty the next processing slice needs to be scheduled (requestIdleCallback(), setTimeout()).
      • when the delivery queue is empty use queueMicrotask() to start the next processing slice when a message is placed on the delivery queue arriving from another node or a general send (e.g. caused by an input event).

Now a system like this is inherently asynchronous—so in my judgement has little to no chance to be adopted by the JavaScript community given how relieved everybody was when async/await was introduced.

In terms of the PolymerLabs MessageStore I would be concerned that forcing messages through IndexedDB would slow things down unnecessarily.

Is postMessage slow?

But thanks for asking …

---

it would be nice with Niladic Components (no props)

Note what is going on in his niladic solution:

He is failing to extract the variation of two near identical components in order to preserve their "niladic" property

import './styles.css';
import { useEffect, useState, useCallback } from 'react';
import { external } from './External';

function Value({ update }) {
  const [value, setValue] = useState();
  const onChange = useCallback(
    ({ target: { value } }) => {
      setValue(value);
      update(value);
      external.doFinalCall();
    },
    [update]
  );

  return (
    <span>
      <input value={value} onChange={onChange} />
    </span>
  );
}

const left = (value) => external.left = value;
const right = (value) => external.right = value;

export default function App() {
  const [total, setTotal] = useState(0);

  useEffect(() => {
    external.registerFinalCall(setTotal);
  }, [total]);

  return (
    <div className="App">
      <h1>Totals</h1>
      <Value update={left} />
      <Value update={right} />
      total is {total}
    </div>
  );
}

props will be involved whenever you use a component more than once on the same page to represent separate business entities.

[3Shain]

props will be involved whenever you use a component more than once on the same page to represent separate business entities.

As I said before

I need to mention that the identity of Component is not always the same as the identity of Instance of Component. That's why in some framework you need to specify the key.

although there are no real key here but different update props are passed to identify different entities. In certain degree it's the mismatch between UI and application proved again.

and the solution is fairly simple... encapsulate High Order Component

although it's not necessary to involve any architecture decision, I'm placing the kairo's solution here

const [External, ExternalImpl] = createConcreteConcern('External', function*() {
    const [leftValue, setLeft] = yield* State(0);
    const [rightValue, setRight] = yield* State(0);
    return {
        leftValue, rightValue,
        setLeft, setRight
    }
}); // a shortcut of `createConcern` with a default implementation

const Value = UI(function*({value, setValue}:{value:Cell<number>, setValue:Setter<number>}) {
    const onChange = e => setValue(e.target.value);
    return Component((_,$)=>(<span><input value={$(value)} onChange={onChange}/></span>));
});

const App = UI(function*() {
    const {leftValue, rightValue, setLeft, setRight} = yield* External;
    const Left = yield* Include(Value.withProps({value:leftValue, setValue:setLeft}));
    const Right = yield* Include(Value.withProps({value:rightValue, setValue:setRight}));
    return Component((_,$)=>(
    <div className="App">
      <h1>Totals</h1>
      <Left />
      <Right />
      total is {$(leftValue) + $(rightValue)}
    </div>));
});

notably I made Value fully controlled because it's unnecessary to have a local state as there is already a source of truth.

[andreyvolokitin]

@peerreynders regarding Actor-Model you mentioned in this thread, you might be interested in this lib: reatom.dev/ (but it also relevant to the overall discussion)