React is an amazing library. Over the last 5 years it has transformed the landscape of frontend development and spawned an incredible ecosystem of tools, libraries, and patterns. Over that time, React has changed quite a bit. As React has evolved, so too have our applications, tools, and approaches.

This year (2019), React went through one of its biggest changes with React v16.8: The One With Hooks. Combined with the official Context API in v16.3, this update dropped a grenade into the React world. As developers like Kent C. Dodds pointed out, complex state management became something that any React developer could implement—without bringing in tools like Redux, MobX, or Apollo.

So we (by we I mean “me”) did what Dan Abramov from the React team said not to do. We started rewriting our app in Hooks…and it was glorious! Hooks helped us pull behavior into common locations, re-use that behavior across other components, and catch prop-state syncing bugs that most of us didn’t know about.

Something else happened. Once we got that behavior pulled up and isolated, we took a look at it. Realization was dawning upon us: Our components knew too much!

What did our components know?

  • Our API backend
  • Our data model
  • Our business logic

What did this mean? We had to update our React components whenever any of those things changed. It laid the groundwork for future issues:

  • Shotgun surgery: one change results in many components needing edited
  • Divergent changes: one component needs multiple edits to accomodate one change
  • Duplicated code: same code structure in multiple components that requires parallel edits for a change
  • Primitive obsession: reluctance to create useful data types which results in repetitive low-level logic
  • Repeated conditionals: same conditional switching logic in different components

With this realization, my team and I started experimenting with patterns to reduce how much our components knew. Through research and experimentation, we discovered the following: a pattern already existed, and you’ve probably heard of it.

Model View Controller (MVC)!

MVC?!

Probably the widest quoted pattern in UI development is Model View Controller (MVC)—it’s also the most misquoted.
  — Martin Fowler

The guiding light of Model View Controller (MVC) is separating presentation from domain. Why is that important to do? Our application’s “domain” is where we model our perception of the problem and its solution. By making this code separate—without reference to any UI—it could be modeled more correctly, tested more deeply, and presented more numerously.

The most important part of MVC is the model. In truth, you aren’t doing MVC until you have a model. Sadly, “model” is a hugely overloaded term (especially in the object-oriented patterns space). In this case, we will define it more similarly to the broader concept of a “data model”: a construct to contain your domain-specific data and logic. Ideally, a model would have no idea a UI even existed. For our purposes, don’t think of a model as “an instance of a class inheriting from a Model object” but rather “a collection of functions and objects that are specific to my app’s domain.”

MVC in React?!

But MVC is an object-oriented programming (OOP) pattern, and React isn’t an object-oriented library, is it? It’s a functional programming (FP) library, right? But all kinds of alternative view libraries popped up after React specifically because React isn’t FP enough (looking at you Cycle.js 😉).

Gahhh!!! What is React?!

Like most code, React isn’t soley OOP or FP; it’s a mix of both and that’s OK. Not only that, but it fits inside a broader application that can be modeled with either approach.

So why MVC in React? As far back as Pete Hunt, React has been described as the “V in MVC”. I haven’t heard anyone describe React that way in years because it doesn’t make as much sense in the current JS landscape. When you put that statement in the context of the times (Ember, Angular 1, Backbone, etc.), what I think Pete was really saying was “React doesn’t tell you how to do models”.

That was a big departure at the time because models—or similar positioned constructs with different names (looking at you Angular)—were a prime feature of all the big frontend frameworks. By leaving this out, React was bucking a trend to instead focus on making a productive view library. It worked!

However, we felt the pain around losing our frameworks on day 1. Facebook told us about Flux Architecture which they turned to after experiencing growing pains with their previous MVC architecture (after which they boldly declared that “MVC doesn’t scale”). At first, they only provided high-level overviews and no code but eventually released a Flux Dispatcher example implementation. In time, after many other libraries implemented the Flux architecture, Dan Abramov gave us Redux. Redux implemented a similar architecture and became somewhat of a standard architectural pattern for React applications.

If you find yourself productive in a Flux/Redux-style architecture, that’s great! No need to rewrite. However, I contend that the “MVC doesn’t scale” argument is overstated and invite you to read on and experiment with some of the following ideas.

Implementing MVC patterns in React

The MVC pattern described here breaks down into the following two pillars:

  1. A Presentation Layer of Controller and View React Components
  2. A UI-Agnostic Data Model

Pillar 1: Presentation layer of controller and view React components

This pillar is about separating components by their role regarding access/knowledge of domain objects and logic. In other words, we’re categorizing components by (a) what they know about and (b) what they can do. We group components into two categories:

  1. Controller Components
  2. View Components

A “controller component” knows a lot about the rest of the world. It knows how to access and update “domain data” (application state) and how to choose and execute “domain logic”. For instance, a controller component may know how to query and mutate data via a RESTful API or read/update objects stored in React Context. Generally, controllers are aware of context, side effects, and domains (application state and behavior).

Contrast that with a “view component”, which should be agnostic of most things a controller would know about. A view component shouldn’t know anything about application state (reading or writing), network protocols, or non-UI providers farther up the chain. Views shouldn’t know what protocol you use to speak to a backend or the format that data takes. Views shouldn’t know about your custom state contexts and providers for sharing domain data (application state).

Views and controllers are both allowed to have their own state, but state in views is only for UI purposes. As such things most often belong in controllers, a view should not call hooks useEffect() and useContext() except for UI-specific cases like the following:

  1. Accessing context for UI-specific data and behavior: e.g. theming and routing
  2. Syncing prop changes with local state with useEffect()

Code example

Here’s an example component that gives us an opportunity to refactor.


function App() {
  return <EditCustomer id={1} />;
}

function EditCustomer({ id }) {
  let { customers, dispatch } = useCustomers(); // access context and probably trigger side effects

  let customer = customers.find(c => c.id === customerId);
  if (!customer) {
    return <NotFound />;
  }

  let [errors, setErrors] = React.useState()
  let [saving, setSaving] = React.useState(false)
  let [name, setName] = React.useState(customer.name);
  let [email, setEmail] = React.useState(customer.email);

  let saveCustomer = () => {
    setSaving(true)
    fetch({
      url: `/api/customers/${id}`,
      method: "PUT",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({ name, email })
    })
      .then(response => response.json())
      .then(apiCustomer => {
        // Formatting for differences between backend and frontend
        //   e.g. Rails/Django snake_case into JavaScript camelCase
        dispatch({
          type: "UPDATE_CUSTOMER",
          payload: formatChangeForFrontend(apiCustomer)
        });
      })
      .catch(error => {
        setErrors(error)
      });
      .finally(() => {
        setSaving(false)
      })
  };

  return (
    <div>
      {errors && <ErrorDisplay errors={errors} />}
      <input
        type="text"
        name="name"
        value={name}
        onChange={e => setName(e.target.value)}
      />
      <input
        type="text"
        name="email"
        value={email}
        onChange={e => setEmail(e.target.value)}
      />
      <button onClick={saveCustomer} disabled={saving}>Save</button>
    </div>
  );
}

Refactor to controller-view pattern

Here we will separate the UI from the domain logic.


function App() {
  return (
    <EditCustomerController id={1} />;
  )
}

// Notice explicit suffix "Controller"
function EditCustomerController({ id }) {
  let { customers, dispatch } = useCustomers();

  let customer = customers.find(c => c.id === id);
  if (!customer) {
    return <NotFound />;
  }

  let onSave = (newCustomerData) => {
    return fetch({
      url: `/api/customers/${id}`,
      method: "PUT",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify(newCustomerData)
    })
      .then(response => response.json())
      .then(apiCustomer => {
        // Formatting for differences between backend and frontend
        //   e.g. Rails/Django snake_case into JavaScript camelCase
        dispatch({
          type: "UPDATE_CUSTOMER",
          payload: formatChangeForFrontend(apiCustomer)
        });
      })
  };

  return <CustomerForm onSave={onSave} initialName={customer.name} initialEmail={customer.email} />
}

// Notice no special name; just a React component that knows about React things
function CustomerForm({onSave, initialName, initialEmail}) {
  let [errors, setErrors] = React.useState()
  let [saving, setSaving] = React.useState(false)
  let [name, setName] = React.useState(initialName);
  let [email, setEmail] = React.useState(initialEmail);

  let onSaveWrapped = () => {
    setSaving(true)
    onSave({name, email})
      .catch((error) => {
        setErrors(error)
      })
      .finally(() => {
        setSaving(false)
      })
  }

  return (
    <div>
      {errors && <ErrorDisplay errors={errors} />}
      <input
        type="text"
        name="name"
        value={name}
        onChange={e => setName(e.target.value)}
      />
      <input
        type="text"
        name="email"
        value={email}
        onChange={e => setEmail(e.target.value)}
      />
      <button onClick={onSaveWrapped} disabled={saving}>Save</button>
    </div>
  );
}

Upsides of this code:

  1. View component <CustomerForm /> is far easier to test than the previous component. It requires no mocking of the network or provider/context setup. Tests of it are unit tests.
  2. Any tests for the original <EditCustomer /> component will still pass after you update the import/name! Those tests are now integration tests.
  3. All of the non-UI concerns are isolated in the controller component <EditCustomerController /> and passed into the view. All the view knows is it gets initial values for name and email and async onSave() callback.
  4. The coupling between the view and the rest of our app is minimal. This view could be dropped into any other location of your React tree.
  5. The UI concerns around form control, error, and saving states is kept inside the view.
  6. This same pattern works in TypeScript as well.

If you’re sticking with JavaScript as your language, you could even further decouple the controller from the view thusly:


function App() {
  return (
    <EditCustomerController id={1}>
      <CustomerForm />
    </EditCustomerController>
  );
}

function EditCustomerController({ id, children }) {
  // ... all the same code
  return React.cloneElement(children, {
    initialName: customer.name,
    initialEmail: customer.email,
    onSave
  });
}

// ... view stays the same

Some of you may feel uneasy, but stick with me for a moment.

Upsides of this code:

  1. Same testing benefits.
  2. The controller is uncoupled from the view and could be composed with any other view that accepts those props.
  3. The composing parent could add additional props to the view that it knows about and the controller may not.

Item #3 is especially interesting in cases like the following where the <App/> had some bit of additional data that the controller was ignorant of:


function App() {
  let importantData = {
    /* things */
  };
  return (
    <EditCustomerController id={1}>
      <CustomerForm importantData={importantData} />
    </EditCustomerController>
  );
}

// ... controller stays the same

function EditCustomerController({ id, children }) {
  // ... all the same code
}

// ... view accepts extra prop `importantData`
function CustomerForm({ onSave, initialName, initialEmail, importantData }) {
  // does things
}

Pillar 2: UI-agnostic data model

Now that our controller and view have been separated, let’s look at our controller:


let { customers, dispatch } = useCustomers();

let customer = customers.find(c => c.id === id);
if (!customer) {
  return <NotFound />;
}

let onSave = newCustomerData => {
  return fetch({
    url: `/api/customers/${id}`,
    method: "PUT",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify(newCustomerData)
  })
    .then(response => response.json())
    .then(apiCustomer => {
      // Formatting for differences between backend and frontend
      //   e.g. Rails/Django snake_case into JavaScript camelCase
      dispatch({
        type: "UPDATE_CUSTOMER",
        payload: formatChangeForFrontend(apiCustomer)
      });
    });
};

If you take away the useCustomers() custom hook, how much of that looks like presentation code? Or stuff that React should care about? Not much? Then …

WHY IS IT IN OUR COMPONENT!?

You know what that looks like? It looks like domain logic! Where does domain logic go? Definitely not in our presentation layer. Why should our React component know that stuff? What kinds of things does it know?

  1. Our data is accessed via fetch at an HTTP URL
  2. Our backend speaks JSON
  3. Our backend gives us data with a non-friendly casing (snakes on our GUI!)
  4. Our backend expects the keys “name” and “email” when creating a customer

And this is just a simple, contrived example. I guarantee you’ll find all kinds of far better domain logic in your actual, production applications.

How could we fix this? Pull it out of our component! Let’s start with something easy, like the handler:


function EditCustomerController(props) {
  // ...

  let onSave = async newCustomerData => {
    // NOTE: new line!
    let latestCustomer = await performCustomerUpdate(props.id, newCustomerData);

    dispatch({ type: "UPDATE_CUSTOMER", payload: latestCustomer });
  };

  // ...
}

// NOTE: new function!
async function performCustomerUpdate(id, newCustomerData) {
  let response = await fetch({
    url: `/api/customers/${id}`,
    method: "PUT",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify(newCustomerData)
  });
  let apiCustomer = await response.json();
  return formatChangeForFrontend(apiCustomer);
}

We just pulled this function out of our React component because it wasn’t presentation logic. There is more logic we could do this with. In this case, we used a simple function, but what if you wanted to do something more complicated?


function EditCustomerController(props) {
  // ...

  let onSave = async newCustomerData => {
    // NOTE: yep, that's a `new` keyword
    let gateway = new CustomerGateway();
    let latestCustomer = await gateway.update(props.id, newCustomerData);

    dispatch({ type: "UPDATE_CUSTOMER", payload: latestCustomer });
  };

  // ...
}

// NOTE: A class!!!
class CustomerGateway {
  constructor(fetchFn = fetch) {
    this.fetch = fetchFn;
  }

  async update(id, data) {
    await this.fetch({
      url: `/api/customers/${id}`,
      method: "PUT",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify(data)
    });

    let formattedData = this.formatChangeForFrontend(await response.json());

    // do more things ...
    return formattedData;
  }

  formatChangeForFrontend(apiData) {
    // return transform data from snake_case to camelCase
  }
}

We could even go another step toward passing these functions/objects in as props to our controller component.


function EditCustomerController({ Gateway = CustomerGateway }) {
  // ...

  let onSave = async newCustomerData => {
    let gateway = new CustomerGateway();
    let latestCustomer = await gateway.update(props.id, newCustomerData);
    // ...
  };

  // ...
}

/* OR */

function EditCustomerController({ updater = performCustomerUpdate }) {
  // ...

  let onSave = async newCustomerData => {
    let latestCustomer = await updater(props.id, newCustomerData);
    // ...
  };

  // ...
}

These contrived examples are meant to illustrate two things:

  1. Keep moving non-UI behavior and knowledge out of components.
  2. Use the modeling technique that (a) you are comfortable with and (b) fits the problem.

Are you great at functional programming? Use it to handle your domain logic. Write pure functions. Compose those functions together. Put your domain data into persistent, immutable structures. You do you!

Are you great with object-oriented programming? Do it! Make classes. Compose those classes together. Practice SOLID principles. You do you!

Are you like most developers and basically comfortable with both? Mix them! Well-designed systems and code, regardless of the “paradigm”, look very similar and trend toward short, precise, low-coupled functions and methods.

None of these things need to seriously impact your React code. In fact, they shouldn’t. Because regardless of how you get there, separating your presentation and domain will help you. Testing will be easier. Changes will be easier to implement. Business changes won’t send you spelunking into your React tree to update every instance of a certain if...else statement that is broken now that your API added a new value to the returned data.

Contrast: Container and presentational components

Container and Presentational Components came from Dan Abramov, and the idea has been important in the Redux community. Controller and View Components are very similar. Generally, Controller == Container && View == Presentational.

In 2019, Dan Abramov added a disclaimer on his post:

I wrote this article a long time ago and my views have since evolved. In particular, I don’t suggest splitting your components like this anymore. If you find it natural in your codebase, this pattern can be handy. But I’ve seen it enforced without any necessity and with almost dogmatic fervor far too many times. The main reason I found it useful was because it let me separate complex stateful logic from other aspects of the component. Hooks let me do the same thing without an arbitrary division. This text is left intact for historical reasons but don’t take it too seriously.
  — Dan Abramov

I think I understand his change of position, but disagree for a few reasons:

  1. Idealistic, dogmatic adherence to a pattern is possible and negative with any pattern. Avoiding such dogma is a discipline unconnected to the patterns themselves.
  2. In practice, most custom hooks still fall into one of two roles: UI hooks and non-UI hooks (e.g. data fetching, mutation, and caching). Custom hooks rarely do both; always one or the other. Calling UI hooks from View components and non-UI hooks from Controller components is simple and provides useful separation.
  3. Regardless of the method used (i.e. hooks or lifecycle methods), explicitly recognizing components by roles is very helpful for testing, organization, and re-use.
  4. Determining where to split components is difficult and patterns that provide guidance for that are very helpful. Generally, React developers struggle with having components that are too large and do too much rather than the inverse.

Conclusion

Today, MVC is commonly viewed as a “server-side architecture” that doesn’t map well to GUI programs. This assumption ignores (a) the origins of MVC as a Smalltalk invention specifically for GUIs and (b) the numerous successful implementations of MVC and derivatives (MVVM, MVPM, MVP, MVT) across many UI platforms today (e.g., iOS on mobile, Qt on desktop, Backbone/Angular/Ember on web). MVC is an even better design philosophy for GUIs than it is for server-side systems.

Flux and Redux have been good patterns for the React ecosystem not least because of the exposure to functional programming principles (especially Redux and its influence from Elm). Before React and Redux, FP was primarily a niche concept in web frontends. Today, it’s clearly a preferred pattern in building React applications. Sadly, I worry that the pendulum has swung too far towards FP and caused a narrowing of thought where alternative design strategies are never even considered.

The patterns proposed here, (1) Controller + View Components and (2) UI-Agnostic Data Model, do not require strict OOP or FP styles. This is a good thing because JavaScript is a very general language that favors a mixed style far better than a single one. The patterns also mesh well with the spirit of React as “just a view library”.

I hope you find useful concepts in this proposal that will enable you to manage the complexity of your applications and prevent some of the maintenance pains that plague so many systems.

Tommy Groshong

Person An icon of a human figure Status
Sleeper Agent
Hash An icon of a hash sign Code Name
Agent 0057
Location An icon of a map marker Location
Ohio