krawaller


This post details a refactoring of the previous React case study. We also touch on some Underscore usage.

A React.js case study follow-up

Tags: reactunderscorecase study

React Memory Game version 2

A week ago I wrote a case study on a small React experiment, a memory game I whipped up to learn the ins and outs of React. Here's the game in an iframe again (here's a link if you want it in a separate tab), to save you from hopping back to refresh your memory (...). The repo, as before, can be found here.

Since this was my first foray into React the code was, well, less than perfect. But I was fortunate to receive lots of excellent feedback in the comments and on Hacker News, foremost of which was Facebook's Ian Obermiller who took the time to do a fork of the repo with in-depth comments to each commit.

After digesting all the feedback, and having spent more time with the React docs, I went over the Memory game code again. This post walks through those changes and the lessons learned. The text assumes you've already read the previous post, so please hop over there if you came straight here.

ES6 Transformation

First off, Ian and others pointed out that the JSX Transformation also contains support for some ES6 features, so there's really no reason not to use them! Throughout the new code I've therefore made use of these, primarily the Arrow Functions and Method Definition Shorthand.

The Arrow Function syntax has many nuances, but primarily it is a way to define a function more succinctly. The created function is also bound to the current scope.

// old way
var myFunc = function(arg) {
// do stuff
}.bind(this);

// ES6 way
var myFunc = arg => {
// do stuff
};

And here's the method definition shorthand example:

// old way
var o = {
method: function() {
return "Hello!";
}
};

//ES6 way
var o = {
method() {
return "Hello!";
}
};

The new Game component code

Now to the React stuff! Starting again with the top-level Game component, here's the updated code:

/** @jsx React.DOM */

var Game = React.createClass({
getInitialState() {
return {};
},
startGame(words) {
this.setState({
words: _.shuffle(words.concat(words))
});
},
endGame() {
this.setState({ words: undefined });
},
render() {
return this.state.words ? (
<Board onEndGame={this.endGame} words={this.state.words} />
) : (
<Wordform onWordsEntered={this.startGame} />
);
}
});
PropsStateSub componentsInstance variables
playing
tiles
Wordform
Board

For each component I'll also show the same table as before, highlighting any differences.

As you can see here, I removed the playing flag from Game, opting instead to set words to undefined on game end. Slightly less obvious code, but one less state variable. Which is preferrable is pure philosophy, and mine is to go for brevity.

Show & hide versus rendering only relevant components

in my first version I alternated between showing the Board and Wordform through always rendering both, but showing and hiding them as appropriate. This is the old render function containing this approach:

return (
<div>
<div className={this.state.playing ? "showing" : "hidden"}>
<Board
endGame={this.endGame}
tiles={this.state.tiles}
max={this.state.tiles.length / 2}
key={this.state.seed}
/>

</div>
<div className={this.state.playing ? "hidden" : "showing"}>
<Wordform startGame={this.startGame} />
</div>
</div>

);

By misreading the docs and misunderstanding an error message, I was fooled into believing that this was the way to go, instead of choosing what component to render. As it turns out, there is no problem with the latter approach, which of course makes for much cleaner code:

return this.state.words ? (
<Board onEndGame={this.endGame} words={this.state.words} />
) : (
<Wordform onWordsEntered={this.startGame} />
);

Note also how we no longer need to throw a random seed into the key property of Board to force it to rerender, as it will always be a new board since it previously didn't exist in the virtual DOM.

A side note; if we used React with addons, we could use ReactCSSTransitionGroup to animate the components as they enter/leave.

Decoupling through callback names

Sharp eyes will also note how the new version passes this.startGame to Wordform as a property called onWordsEntered instead of startGame. This was another good point of feedback from Ian; name your properties to make the child less coupled to the parent. Wordform doesn't care what happens when the words are submitted, it merely calls a callback when that happens. The name should reflect that.

Similar renamings have been done throughout the code base.

The new Wordform component code

Here's what the new Wordform component code looks like:

/** @jsx React.DOM */

var Wordform = React.createClass({
propTypes: {
onWordsEntered: React.PropTypes.func.isRequired
},
getInitialState() {
return { error: "" };
},
setError(msg) {
this.setState({ error: msg });
setTimeout(() => {
this.setState({ error: "" });
}, 2000);
},
submitWords(e) {
var node = this.refs["wordfield"].getDOMNode(),
words = (node.value || "")
.trim()
.replace(/\W+/g, " ")
.split(" ");
if (words.length <= 2) {
this.setError("Enter at least 3 words!");
} else if (words.length !== _.unique(words).length) {
this.setError("Don't enter duplicate words!");
} else if (_.find(words, w => w.length > 8)) {
this.setError("Words should not be longer than 8 characters!");
} else {
this.props.onWordsEntered(words);
node.value = "";
}
return false;
},
render() {
return (
<form onSubmit={this.submitWords}>
<p>Enter words separated by spaces!</p>
<input type="text" ref="wordfield" />
<button type="submit">Start!</button>
<p className="error" ref="errormsg">
{this.state.error}
</p>
</form>

);
}
});
PropsStateSub componentsInstance variables
*onWordsEntered()*error

The only change in the table is the already mentioned renaming, shown here through the italizicing of onWordsEntered.

Defining expected properties through propTypes

The primary point of interest in Wordform is up top; Ian made me aware of the propTypes property which allows you to clearly communicate what properties your component expects (and also validate them while in development mode).

propTypes: {
onWordsEntered: React.PropTypes.func.isRequired
},

As we see here Wordform just expects the one, namely the previously mentioned callback which got a new name.

Semi-defining the state variables

Note our call to getInitialState:

getInitialState(){
return {error:''};
}

Returning an empty object here would suffice, but explicitly setting the error property to an empty string communicates that the component will make use of this state variable.

This, along with the propTypes literal above, goes a long way to help the reader to gain immediate understanding of the component's functionality. Together they give you the same information as my data table!

Underscore elitism

There is one other tiny change regarding Wordform; in the old code, I used the following expression to test if any word was too long;

_.filter(words, function(w) {
return w.length > 8;
}).length;

...while in the new code, I'm doing this (shown here without ES6 stuff):

_.find(words, function(w) {
return w.length > 8;
});

Functionally it makes absolutely no difference, but it touches on something I feel is important, albeight on a pedantic level - knowing your tools is important. Learning Underscore/Lodash (along with functional programming) levelled me up quite a bit as a programmer, and I took pride in being able to write shorter code. Making this "mistake" therefore itched quite a bit.

The new Board component code

Now for the big one! Board was already the most complex component, so it is only natural that it contains the most changes. Here's the new code in full:

/** @jsx React.DOM */

var Board = React.createClass({
propTypes: {
words: React.PropTypes.arrayOf(React.PropTypes.string).isRequired,
onEndGame: React.PropTypes.func.isRequired
},
getInitialState() {
return {
found: 0,
message: "choosetile",
tilestates: _.map(_.range(this.props.words.length), () => "unturned")
};
},
componentWillMount() {
this.max = this.props.words.length / 2;
},
clickedTile(index) {
if (this.state.tilestates[index] === "unturned") {
// turn up lone tile
if (this.flippedTileIndex === undefined) {
this.flippedTileIndex = index;
this.setState({
message: "findmate",
tilestates: _.extend(
this.state.tilestates,
_.object([index], ["revealed"])
)
});
// clicked second tile
} else {
var otherindex = this.flippedTileIndex,
matched = this.props.words[index] === this.props.words[otherindex];
delete this.flippedTileIndex;
// found mathing pair
if (matched) {
this.setState({
found: this.state.found + 1,
message: "foundmate",
tilestates: _.extend(
this.state.tilestates,
_.object([index, otherindex], ["correct", "correct"])
)
});
// pair didn't match
} else {
this.setState({
message: "wrong",
tilestates: _.extend(
this.state.tilestates,
_.object([index, otherindex], ["wrong", "wrong"])
)
});
}
// restore UI message after 1500, and flip back eventual failed attempt
setTimeout(() => {
if (this.isMounted()) {
this.setState({
message:
this.state.message === "findmate"
? "findmate"
: this.max === this.state.found
? "foundall"
: "choosetile",
tilestates: matched
? this.state.tilestates
: _.extend(
this.state.tilestates,
_.object([index, otherindex], ["unturned", "unturned"])
)
});
}
}, 1500);
}
}
},
render() {
return (
<div>
<button onClick={this.props.onEndGame}>End game</button>
<Status
found={this.state.found}
max={this.max}
message={this.state.message}
/>

{this.props.words.map((b, n) => (
<div onClick={_.partial(this.clickedTile, n)}>
<Tile word={b} status={this.state.tilestates[n]} />
</div>
))}
</div>

);
}
});
PropsStateSub componentsInstance variables
*words*
*onEndGame()*
found
message
**tilestates**
Status
Tile
wait
*flippedTileIndex*
**max**

As said, lots of changes, which we'll now walk through!

Waiting is boring

First off, the old version contained a wait flag which made sure that after you turned up a pair, you couldn't click a new tile for 2 seconds while the succeed/fail animations took place. This "feature" was badly communicated and rather frustrating, and most people thought it was a bug.

Therefore I decided to simply remove it, hence the stricken out wait instance variable in the component data table.

Property-generated constant

In the previous version Board didn't make use of max. It was just passed as a property to Status, calculated at that point:

<Status
found={this.state.found}
max={this.props.tiles.length / 2}
message={this.state.message}
/>

Inside Status we used the max information to decide what to display when the message instruction from Board was choosetile:

this.props.message === "choosetile" && found === max
? "foundall"
: this.props.message;

Going over the code it felt weird to have just that particular logic in Status. It seemed to belong in the parent, so I decided to move it there. But that means Board now needs to use the max value, which raised a seemingly innocent question - how to handle that?

The max value is calculated by doing words.length/2, but I don't want to calculate it on the fly every time I need it. An older version of the code passed max along with words from Game to Board, but that seems silly too. It should be calculated from words when the Board instance is initialized.

The premier place for turned out to be the componentWillMount hook. At this point in the component life cycle the properties are accessible, so max can be set as an instance variable like thus:

componentWillMount() {
this.max = this.props.words.length/2;
},

Throughout the rest of the code I could then replace this.props.max with this.max. A small gain, but the real win was cleaning up the logic in the child component, Status, without having to increase the complexity of Board's signature by expecting max as a property.

Note how the componentWillMount call sits up top with propTypes and getInitialState, as it also is communicating contents from my component data table!

State variable VS instance variable

I noticed Ian refactored the old wait instance variable to be a state variable, but I can't (yet?) see the advantage to that. Why put variables in this.state, with all the overhead that means, if they will never affect the output of the render method?

That's why I didn't make max a state variable, which would of course also have worked:

getInitialState() {
return {
max: this.props.words.length/2,
// the other variables redacted
};
},

It would even mean less code, as the one extra line in the getInitialState object literal replaced the call to componentWillMount. Still, in my head, using this.state solely for rendering-affecting state took precedence.

The new Status component code

Here's the full code for the updated Status component.

/** @jsx React.DOM */

var Status = React.createClass({
propTypes: {
found: React.PropTypes.number.isRequired,
max: React.PropTypes.number.isRequired,
message: React.PropTypes.oneOf([
"choosetile",
"findmate",
"wrong",
"foundmate",
"foundall"
]).isRequired
},
render() {
var found = this.props.found,
max = this.props.max,
texts = {
choosetile: "Choose a tile!",
findmate: "Now try to find the matching tile!",
wrong: "Sorry, those didn't match!",
foundmate: "Yey, they matched!",
foundall: "You've found all " + max + " pairs! Well done!"
};
return (
<p>
({found}/{max})&nbsp;&nbsp;{texts[this.props.message]}
</p>

);
}
});
PropsStateSub componentsInstance variables
found
max
message

The only change is the already discussed extraction of logic deciding whether to show foundall or choosetile.

Note also the neat syntax for defining an enum proptype, used here for the message property.

Performance-wise it would probably be beneficial to move the texts variable out of the createClass call, to prevent it having to be created everytime render is called. As is, though, it hardly matters.

The state of a tile

In the old version, the Tile component made use of three state variables; flipped, correct and wrong. This was really bad design on my part, which also made for needlessly complex code. When a tile is flipped, it is either correct or wrong. If wrong, I'll reset flipped to false after 2 seconds.

A much better design is to bake this into a single tile status variable with the possible values of unturned, correct and wrong!

So why are we discussing the innards of Tile while going over changes to Board? Because having refactored everything into that single status enum variable, I realised that this value should be passed to Tile from Board. Before, when they were three different variables, I didn't even see the possibility. Ian did, so his new version of Board contains two state arrays tracking this; wrongIndexes and correctIndexes.

Having baked it all together into an enum variable, my approach was instead to give Board a single tilestates array containing these values. We initially set them all to unturned:

getInitialState() {
return {
tilestates: _.map(_.range(this.props.words.length),()=>'unturned'),
// the other variables redacted
};
},

When we render the tiles, each tile is passed its status as a property along with the word:

<Tile word={b} status={this.state.tilestates[n]} />

As the game progresses, we merely need to update the corresponding indexes in tilestates in Board, and the faux data binding of React's "rerender everything" approach will take care of the rest!

More underscore shenanigans

Speaking of updating the tilestates array, here's a closer look at the code where that is done:

// revealing a lone tile
this.setState({
message: "findmate",
tilestates: _.extend(this.state.tilestates, _.object([index], ["revealed"]))
});

// marking a pair as correct
this.setState({
found: this.state.found + 1,
message: "foundmate",
tilestates: _.extend(
this.state.tilestates,
_.object([index, otherindex], ["correct", "correct"])
)
});

// marking a pair as wrong
this.setState({
message: "wrong",
tilestates: _.extend(
this.state.tilestates,
_.object([index, otherindex], ["wrong", "wrong"])
)
});

// turning a pair back down
this.setState({
message:
this.state.message === "findmate"
? "findmate"
: this.max === this.state.found
? "foundall"
: "choosetile",
tilestates: matched
? this.state.tilestates
: _.extend(
this.state.tilestates,
_.object([index, otherindex], ["unturned", "unturned"])
)
});

The _.extend call calculating the new tilestates array is succinct to the point of arrogance, but I feel it is warranted here. Expressing this functionality as a one-liner makes the code much less verbose, which I argue is worth the heightened cognitive cost of understanding what the hell the line actually does. This is understood contextually, even for those who can't decipher the code.

There's another reason why I didn't make a helper function containing the same functionality in a more readable way; we should never mutate anything in this.state except through calls to this.setState. And not mutating an array while we're operating on it over several LOC's isn't feasible. That means we'd have to copy the array, mutate the copy, and then finally pass that to setState.

Doing it that way would amount to lots of work and lots of LOC's, which I use as an excuse to get away with my arrogant one-liner. I concede that there are probably times when I make this exact argument that I'm in the wrong, but the point remains; wielded with responsibility, the powerful expressiveness of Underscore/Lodash can really help you make the code less bulky.

Tile click catching attempt #1 - passing back instance

When a Tile is clicked, we need to act on this in Board. In my first version, I pass a click handler from Board to Tile. Inside tile, the tile will then call the handler passing itself as a parameter.

// old click handler inside `Tile`
catchClick: function(){
if (!this.state.flipped){
this.props.clickedTile(this);
}
},

This way board got access to the instance, and can call the relevant methods on the tile to show or hide it.

// old clickedTile method in Board was passed tile instance and called methods on that
clickedTile: function(tile){
if (!this.wait){
if (!this.flippedtile){
tile.reveal();
// ...rest redacted...
} else {
this.wait = true;
if (this.flippedtile.props.word === tile.props.word){
tile.succeed();
this.flippedtile.succeed();
// ...rest redacted...
} else {
tile.fail();
this.flippedtile.fail();
// ...rest redacted...
}
// ...rest redacted...
}
}
},

This is needlessly complex, and, as Ian pointed out, an antipattern when React structure is concerned. You should never pass a component instance upstream, there's always a better choice.

Tile click catching attempt #2 - passing back index

After refactoring the tile status to the tilestates array in Board, it would be enough if we told Board the index of the clicked Tile. In this version the rendering of the tiles inside the Board render method looked like this:

{
this.props.words.map(function(w, n) {
return (
<Tile
clickedTile={this.clickedTile}
word={w}
status={this.state.tilestates[n]}
key={n}
/>

);
});
}

And in Tile, the click handler passes the key property instead of the instance as in attempt #1.

catchClick: function(){
if (this.props.status==='unturned'){
this.props.clickedTile(this.props.key);
}
},

Tile click catching attempt #3 - calling prefilled callback

In attempt #2, we're passing key to Tile only to have Tile pass it right back in the click handler. That seems needlessly clunky. There's no computation done inside Tile that Board is interested in, we're just getting back stuff we already know.

But we have to pass a callback to Tile, and we have to know the index of the clicked tile. That gave me the idea for attempt #3 - how about, instead of passing clickedTile, we pass a function wich calls clickedTile with the correct index? Here's the new render loop doing that, using _.partial to create the callback:

{this.props.words.map(
function(w,n){
return (<Tile clickedTile={_.partial(this.clickedTile,n))} word={w} status={this.state.tilestates[n]} key={n}/>);
}
)}

And here's the new catchClick in Tile, which now just calls the callback:

catchClick: function(){
if (this.props.status==='unturned'){
this.props.clickedTile();
}
},

Tile click catching attempt #4 - moving the unturned check up to Board

Having the check to prevent clicking turned tiles inside Tile didn't really sit right with me. All other computations regarding the status of a tile is right there in Board's clickedTile method. I therefore moved the check here:

clickedTile(index){
if (this.state.tilestates[index]==='unturned'){
// ...redacted ...
}
}

The click handler inside Tile is now reduced to this:

catchClick: function(){
this.props.clickedTile();
},

Tile click catching final version - catching click in Board

It's getting cleaner, but even this sleek version felt weird. Why bother catching the click event inside Tile at all? There's absolutely no computations going on, and we're not passing anything back. Why not simply catch the clicks in Board?

I decided to try this, wrapping all tiles in a div which container the click handler. Here's the final version of the render loop:

{this.props.words.map(
(b,n) => {
return (<div onClick={_.partial(this.clickedTile,n)}>
<Tile word={b} status={this.state.tilestates[n]} />
<div>)
}
)}

There's an added cost in form of the wrapping div tags, but the upside is that Tile no longer contains a click handler at all.

Whether this final version is actually cleaner than attempt #4 is definitely debatable. As clickedTile is already very complex, you could even argue that attempt #3, where the unturned check lives in the Tile click handler, is the ideal solution.

Having the click handler inside Tile also lets you put it on the clickable side. Here's the render code for Tile before this final version:

render(){
return (
<div className={'brick '+this.props.status}>
<div className="front" onClick={this.catchClick}>?</div>
<div className="back">{this.props.word}</div>
</div>
);
}

However, the value of the "correct" placement of the click handler is limited - since we're animating the flipping of tiles, we still need to check the status of a clicked tile, as it would otherwise be possible to do a quick doubleclick on a tile and match it to itself.

The new Tile component code

Finally, here's what's left of the now pitiful Tile component:

/** @jsx React.DOM */

var Tile = React.createClass({
propTypes: {
status: React.PropTypes.string.isRequired,
word: React.PropTypes.string.isRequired
},
render(){
return (
<div className={'brick '+this.props.status}>
<div className='front'>?</div>
<div className='back'>{this.props.word}</div>
</div>
);
}
});
PropsStateSub componentsInstance variables
word
clickedTile()
**status**
flipped
wrong
correct

Because of my status refactoring I no longer need the css class shenanigans of the previous version, I merely add the value of this.props.status as a class. But if I had kept the previous approach of separate flag variables, I would be better of using the classSet addon. See Ian's version for what that might look like.

Wrapping up (again)

Diving back down, and diving this deep into sometimes trivial details, was again a powerful learning experience for me! I am hugely grateful to Ian for supplying his fork, which really helped me propelling my understanding along.

I hope you poor souls who survived through all these ramblings also managed to get something out of it!

And again, the main takeaway for me was: React is really, really powerful stuff. Absolutely loving it, and can't wait to explore the flux architecture and routing solutions!