One post tagged with "concept"

A previous article discussed how StyleX was initially created to address the needs that arose during the facebook.com rewrite. However, early on, StyleX looked quite different from what it is today.

StyleX was developed to replace our previous styling solution, which was similar to CSS modules. The biggest challenge we faced was the massive amount of CSS and the style recalculation costs of lazy loading styles. So, the very first version of StyleX generated atomic classes.

In fact, that was all StyleX did for some time. We used utilities like classnames to compose styles just like we did with our previous styling system. The API looked quite different back then:

const styles = stylex({
  root: {
    color: 'white',
    backgroundColor: 'blue',
  }
});

<div className={styles('root')} />

It was impossible to use styles outside of the file they were defined in.

The Need for Style Composition​

Initially, it seemed like style composition was unnecessary. After all, we'd been able to get by with classnames to compose styles from CSS modules for years. But as we built a component library on top of StyleX, we realized that we needed style composition. We needed the ability to pass styles around as props to components. It became clear that style composition was already pervasive in our codebase. However, style composition with classnames was inconsistent and unpredictable, and we had learned to accept it.

Style composition has never been easy with CSS. Applying multiple CSS rules to the same HTML element usually leads to a constant battle with CSS order, selector specificity, and growing CSS files. CSS has evolved to provide additional tools such as @layer, which makes the problem slightly better. However, the problem persists.

In any design system, it is common to modify or override certain styles of components within certain contexts. Component libraries need to be customizable to fit different application designs. Without proper style composition, we end up with unnecessary duplication of styles and code.

The rise of runtime CSS-in-JS libraries was fueled, in part, by this need to solve the style composition problem. By being able to generate and inject styles at runtime, these libraries provided a way to compose styles in a way that was not possible with traditional CSS. However, this also came with performance trade-offs.

Learning from Inline Styles​

While exploring solutions, we made an interesting observation: inline styles have always been naturally composable. In HTML, style accepts a string with a list of styles. In this string, the last style applied always wins.

Inline styles have their own limitations, too. They don't support pseudo-classes, media queries, or other selectors. It is also difficult to enforce consistency when using inline styles directly in HTML. However, component-based frameworks, such as React, largely sidestep any architectural issues with inline styles. Components are a layer of abstraction that enables code reuse without needing to write the same styles over and over again. This change was noticed early and was described in the original CSS-in-JS talk by Christopher Cheadeu.

So, when it came to designing StyleX, we decided to model our styles after inline styles. To form a mental model, it can be helpful to think of StyleX as "inline styles without the limitations".

Key Design Decisions​

Static Style Definitions​

By the time we made this conscious decision, StyleX had already been in development for a while and had evolved organically, inspired by React Native's StyleSheet API, which was itself inspired by inline styles.

One of the first design decisions we reconsidered was the requirement to declare stylex.create before using it, without allowing the definition of styles inline. We realized that we had to allow the ability to statically define styles as JavaScript constants and be able to reuse them across multiple elements, multiple components, and even multiple files. Once we had this realization, we felt more comfortable not offering the ability to define styles inline. Even if occasionally inconvenient, it is better to have one consistent way to define styles.

Pseudo-Classes and Media Queries​

Our next design question was to decide how to handle pseudo-classes, media queries, and other at-rules in a way that felt like inline styles and enabled composability. Inline styles don't support pseudo-classes or media queries, but it's possible to use JavaScript to read such 'conditions' and apply styles accordingly:

<div style={{
  color: isActive ? 'white'
    : isHovered ? 'blue'
    : 'black',
  width: window.matchMedia('(min-width: 1200px)').matches ? '25%'
    : window.matchMedia('(min-width: 600px)').matches ? '50%'
    : '100%'
}} />

This approach has obvious performance implications, but it also has some strong architectural benefits. The style object remains flat, making the composition of styles more predictable. We avoid having to deal with complex rules for merging styles and dealing with specificity issues. We don't have to think about the 'default color' or the 'hover color'. We just think about a single 'color' value that changes based on the conditions.

All values for a property being co-located can also lead to a more consistent design system. Instead of thinking about mobile styles or desktop styles, this approach forces you to think responsively about the value of each property.

This realization led to one of our most unique design decisions. Instead of separating "base" styles and styles under a media query, which is common in almost every other CSS library, we decided to treat pseudo-classes and media queries as 'conditions' within the value of a property:

const styles = stylex.create({
  color: {
    default: 'black',
    ':hover': 'blue',
    ':active': 'white',
  },
  width: {
    default: '100%',
    '@media(min-width: 600px)': '50%',
    '@media(min-width: 1200px)': '25%',
  }
});

We found a way to take the concept of JavaScript conditions from inline styles and express them declaratively and statically in a way that can be compiled to static CSS while keeping the architectural benefits.

CSS Shorthands​

CSS shorthand properties introduce another factor of confusion when composing styles. This decision has remained controversial even within the team itself, and so we still support two different strategies for merging shorthands.

React Native pioneered the approach of merging purely by key and assigning different priorities to different properties. Longhand properties have a higher priority than shorthand properties. So, while you can end up with a style object that has both padding and paddingTop, paddingTop will take precedence regardless of the order in which they are applied.

Inline styles, on the other hand, always merge by order of application, and the last style applied always wins. Modeling this behavior in a library that compiles to static CSS wasn't obvious at first. We started by expanding shorthands into longhand properties at compile time. This approach had many edge cases and couldn't be done reliably in all cases.

We eventually discovered a strategy where a shorthand property would also set its constituent longhand properties to null. This would, in practice, "unset" any longhand that may have been applied before. This achieves the exact same behavior as inline styles, where the last style applied always wins. While this approach can result in larger JavaScript objects after compilation, we decided to make it the default behavior as it aligns with our inline styles mental model.

Dynamic Styles​

Finally, we needed to handle dynamic styles. Any API modeled after inline styles must support them, but we knew we needed to do it with care and intention. We wanted to make it possible to use dynamic styles when needed, but we also wanted to make it explicit when styles were dynamic. We don't want an API that makes it easy to accidentally create dynamic styles.

StyleX allows dynamic styles by using functions. Instead of mixing inline objects with static styles created with stylex.create, functions let us define all kinds of styles in a single consistent way.

Behind the scenes, all styles, even dynamic ones, are compiled to static CSS. We don't apply any style property as an inline style. Only CSS variables are ever applied as inline styles and are used as the vehicle to let static styles have dynamic values. By not having any style properties as inline styles, we avoid any conflicts that may arise from inline styles having higher specificity than static styles. (It also makes certain compile-time optimizations possible, but that's its own story.)

Conclusion​

CSS has been around for a long time now. It has evolved in many ways and is now both flexible and one of the most powerful layout models that exist. Yet, it remains a tool with sharp edges and can be challenging to wield effectively.

We've seen many tools and libraries that have tried to make CSS easier to work with. Over the years, many problems have been solved, but the problem of style composition has persisted to some extent. Even when it's possible to compose styles, there have always been inconsistencies and unpredictable behavior.

Yes, we are building a styling solution that is fast, scalable, and easy to use. But we are also building a solution that actually provides predictable style composition without any nasty surprises.