Mehul Kar

August 28, 2016

UI architecture


UI architecture seems to be consolidating around components and some sort of state manager.

React’s whole schtick is about declarative components. Ember’s got the Ember.Component class that smashes together the View and Controller layer. Backbone Views kind of had the same idea, but the name implies larger scope objects. Angular 1x doesn’t really have UI encapsulation, but controllers and $scope could be used to group and isolate UI, I guess.

The point is, encapsulating the behavior of visual elements into objects that directly represent visual elements is the in thing these days. The HTML Web components spec is an even lower level API that captures the same sentiment.

The necessary complement for component architecture to work is an orchestration layer that ties them together and knows what context they live in–i.e. ‘state’. Examples of state are Redux in React world and Services in Ember world. Point is, if you have a set of components plopped onto a page, you need something to tie them togther. Before SPAs, server side provided this functionality.

I like the general direction this is going, but there are a couple open questions with this architecture. (To be clear, these questions haven’t been well answered before this architecture either.)

First, how closely should components be tied to application domain?

The nicest components to write are the lower level ones that serve as either improvements to what’s available in native HTML (e.g. a prettier select box) or additions to native HTML (e.g. toggle UI). These components are nice because they are, by nature, simple, encapsulated, and idempotent; pass in some inputs and get the same behaviour out. Some people compare them to pure functions, I like to think of them as microservices1.

But then there is this other class of components that are more tightly coupled to your application’s domain. An app-header or app-sidebar component, for example. Or a group of the previously mentioned low level components that operate together in a certain way, so it makes sense to create a wrapper component.

This second type of component is a little harder to reason about. It’s hard to design these components, because it’s never clear what state belongs inside and out. But they’re sometimes easier to throw together in a first iteration2.

The point here is that using the ‘component’ moniker for both of these types of objects/UI is a little confusing, both when writing applications and when talking about frontend architecture.

The other overarching thing I’m missing in this architecture is styling. The two problems are that (1) components do not include a spec for styling and (2) there isn’t a clear separation of styles that define the component and styles that are changeable attributes.

So first, I’ve seen that JSX (which is used in React components to inline HTML) can define inline styles. This is a cool concept, but it feels dirty. Maye I’ll come around to it. As I see it, inline styles have these pros and cons:


  • Completes the encapsulation of a UI component
  • Already possible


  • Feels dirty to put ALL style inside the HTML style attribute.
  • Shareability of styles may suffer. This is mostly a nonproblem since JSX can take advantage of ES6 imports/exports, but outside JSX, it would still require a solution.

But putting aside the JSX approach of inline styles, another approach that is exciting would be HTTP 2.0 + CSS Modules. In other words, a component could specify a CSS module that is fetched on render, and with an HTTP 2.0 connection, lots of components could fetch their styles in the same round trip on the network, thus getting scoped styles without performance loss.

The second problem with components and styling is that it’s unclear which styles define the component, and which styles are attributes of the component. For example, if I changed the background color of a Button element, does it become a new component, or is it still a Button that looks different? This answer to this question is unclear in practice and can cause a lot of architectural design pains (and therefore complexity).

Even theoretically, I can’t think of a great answer to this, other than a spec that could separate style definitions between these categories and leaves it to the developer to define their own conventions.

So to recap, frontend programming on the web is maturing. We’ve figured out that thinking about UI in terms of components injected with stateful services is the way to go and that app state doesn’t belong in components.

But we haven’t figured out some of the nuances of this approach. I’m guessing there is some inspiration to be drawn from other GUI paradigms–this is not a new problem (or is it?).

Either way, I still love working on the Web platform. It’s broken and exciting at the same time and I’m glad for both.


  1. Microservice architecture is really just an expression of a pure functions, by the way–at least computational microservices.
  2. Sometimes also a good way to contain technical debt, which is a topic I could write a whole new post on.