Skip to content

Instantly share code, notes, and snippets.

@ChrisShank
Last active May 2, 2024 19:39
Show Gist options
  • Save ChrisShank/7162df4cf8a82112518329165411ce7f to your computer and use it in GitHub Desktop.
Save ChrisShank/7162df4cf8a82112518329165411ce7f to your computer and use it in GitHub Desktop.
Musings on a behavioral paradigm for building web applications

Herein outlines a vision for what a behavioral paradigm for building highly interactive, client-side heavy web applications could look like. It attempts to address problems and challenges with the component paradigm that is commonplace these days. Such challenges include:

  • Component frameworks cluster on templates. Behavior/state is tighly colocated to a particular component in the tree and as requirements change this behavior must be refactored as it's needed in other places. Over time, more and more state is hoisted up to the root of the component tree.
  • Event listeners are directly attached to UI elements. This encourages the event-action paradigm, where event listeners have a increasing amount of conditional logic because it has to figure out the qualatative state the application is in to know what side effects to execute. This leads to a large source of bugs (see Horrock's book "Constructing the USer Interface with Statecharts" for a more elaborate explanation).
  • Behavior is scattered between components. By default there is no coherent way to understand what side effects will be performed as a result of recieving an event in a given component. In the worst case, you have to trace through the entire component tree to find out.
  • State that already exists in the DOM (e.g. input values, focus, etc.) has to be duplicated in the component's JS and synced back to the DOM. Additionally, it's not common to project state back into the DOM via attributes meaning conditional rendering and styling is done programatically in templates instead odf declaratively in CSS.
  • Isomorphic JS means we can usually only use the least common denominator of Web APIs that also work on a server. This excludes undervalued web APIs like DOM, service workers, web workers, audio worklets, paint worklets, background fetch, etc. Some many great features of the web wont be used.
  • The firm encasupation of components make them hard to extend them without bloating the surface API of props/events/slots.
  • Declarative templating systems require hydration which has high overhead and notably a little fragile (e.g. hydration mismatches).

At a high level, here is how a behavioral paradigm would solve these challenges:

  • Clustering on behavior creates a natural shearing layer between the behavior and the UI of an application. These two things likely change at different paces (at least in my own expereince).
  • Use notations like state machines/statecharts to model the core behavior of the application. These paradigms will prevent the bug-riddled "event-action" approach and provide a deterministic and structured way to understand "when the app is in this state and we receive this event, what should happen". Behavior trees and behavior threads could be other notations helpful here.
  • Leverage event bubbling (also know as event delegation) to decouple behavior (how you respond to an event) from the UI (where the event originates from). This makes it easy to progressively enhance HTML and has performance and memory benefits around intialization and hydration.
    • That said, dont rely on DOM structure to understand what a DOM event was meant to do. Instead use HTML attributes to map raw DOM events to their semantic intention. (e.g. <button on:click="INCREMENT">+</button>)
    • Longer explanation here.
    • Repo with prototypes + examples
  • Use custom elements (if possible without the shadowDOM) to progressively enhance parts of the DOM tree in a composable manner. Use attributes on CE's to conditionally style the DOM with no JS and allow more extensible styling. Custom elements are also just classes so they can be easily extended.
  • Routing should be decoupled from the component tree. The "router" should just focus on taking the current state of the application and flattening to into a URL and visa versa. Also maybe let a service worker handle the routing not the main thread.
    • An early prototype of those routing primitives can be found here
  • Be able to move expensive behavior off the main thread (e.g. web worker or service worker). Asynchronus message passing (e.g. actor model or obvservables) is nessecary and can be orchestracted with state machines.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment