Skip to content

Commit

Permalink
Edit pass on the post. More to come
Browse files Browse the repository at this point in the history
  • Loading branch information
@nmn
nmn committed Jan 11, 2025
1 parent 4664c52 commit b4f98cb
Showing 1 changed file with 88 additions and 90 deletions.
178 changes: 88 additions & 90 deletions .../2024-10-10-designing-for-composition.mdx → .../2025-01-10-designing-for-composition.mdx
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,96 +12,99 @@ tags:
- design
---

## Style Composition has never been easy
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.

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.
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.

CSS has evolved to provide additional tools such as `@layer` which makes
the problem slightly better. However, the problem persists and the responsibility
for solving the problem remains on tools such as StyleX.
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:

## Style Composition is desirable
```tsx
const styles = stylex({
root: {
color: 'white',
backgroundColor: 'blue',
}
});

Despite massive challenges with style composition, developers have always
wanted ways to compose styles. Even before CSS-in-JS libraries became
popular, developers used packages such as `classnames` to try and apply
multiple sets of styles to the same element, despite all of the challenges
that come with it.
<div className={styles('root')} />
```

In any design system, it is common to modify or override certain styles
of components within certain contexts. It is also useful to be able to use
component libraries and customize them to fit the design of the application.
The alternative leads to unnecessary duplication of styles and code.
It was impossible to use `styles` outside of the file they were defined in.

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

In recent years, just as we've seen a rise in the popularity of atomic CSS,
we've also seen utilities that merge such styles gain prevalence. Many tools
now offer a way to merge styles.
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.

Despite all the tools, style composition remains a challenge. From headless
UI libraries to design system components that are copied to your own codebase,
we believe many of these tools need to exist entirely because of the inconsistencies
of merging styles.
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.

We believe that we've built the first solution with style composition at its very core.
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.

## Inline Styles have always been composable
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.

Despite the challenges of composing styles with CSS, inline styles have always
been composable. In HTML, `style` accepts a string with a list of styles. In
this string, *the last style applied always wins*.
## Learning from Inline Styles

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.
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*.

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_](https://vimeo.com/116209150).
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_](https://vimeo.com/116209150).

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".

StyleX gives you the ability to use capabilities such as pseudo classes and
media queries, which are not possible with inline styles, all while maintaining
the ability to compose styles in the same way that inline styles do.
## Key Design Decisions

## Do "inline styles" need to be defined inline?
### Static Style Definitions

By the time we made the conscious decision to model StyleX after inline styles,
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.
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.

And so one of the first design decisions we reconsidered was the requirement to
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. There are trade-offs to both approaches, we realised, 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.
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.

The requirement to define styles at the top-level of a module also enforces some of
the requirements of the compiler without feeling *too* unnatural.
### Pseudo-Classes and Media Queries


## Pseudo Classes and Media Queries

Our next design question was to decide how to handle pseudo classes, 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.
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:

```tsx
<div style={{
Expand All @@ -115,20 +118,20 @@ to use JavaScript to read such 'conditions' and apply styles accordingly.
```

This approach has obvious performance implications, but it also has some strong
architectural benefits. The style object remains flat, making composition of
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
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.
the value of a property:

```tsx
const styles = stylex.create({
Expand All @@ -149,44 +152,39 @@ 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 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's Strategy

React Native pioneered the approach of merging purely by key, and assigning
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 Strategy

Inline styles, on the other hand, always merge by order of application, and
the last style applied always wins. Modelling this behaviour in library that
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 the
same behaviour as inline styles, where the last style applied always wins.

This approach has some performance trade-offs. It can result in larger JavaScript objects
after compilation. However, as our overall design philosophy was to
model StyleX after inline styles, we decided to make this the default behaviour.
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
### Dynamic Styles

Any API modelled after inline styles must support dynamic styles, 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.
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
Expand All @@ -203,13 +201,13 @@ certain compile-time optimizations possible, but that's its own story.)

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 it can be challenging to wield effectively.
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 behaviour.
styles, there have always been inconsistencies and unpredictable behavior.

Yes, we are building a styling solution that is fast, scalable and easy to use.
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.

0 comments on commit b4f98cb

Please sign in to comment.