Puzzle: Define HTML custom element subclasses that can fill in base class insertion points

This post presents a little web component architectural puzzle which I’ve come across in the early stages of creating Quetzal. The puzzle deals with how Quetzal should best deliver an important component service on an HTML custom element substrate, and relates specifically to subclassing semantics. Any suggestions or comments would be much appreciated.

Background on the puzzle

Quetzal is an attempt to deliver key features of the QuickUI component model in HTML custom elements. One such feature is that an element subclass should be able to easily populate a slot (insertion points, in HTML parlance) defined by a base class. In practice, there are many situations in which you want to be able to say, “This new UI component should be just like that existing UI component, only with some stuff pre-filled in.” For example:

  • The QuickUI documentation presents a simple page template example in which classes in a small page template hierarchy fill in specific bits of their parents classes.
  • A DateComboBox fills in the popup portion of a ComboBox, which in turn is filling the content portion of a PopupSource. This same facility is also used throughout the QuickUI Catalog. Moreover, it is used in many QuickUI apps in which a stock Catalog component is specialized for the app’s context.
  • Along those same lines, this same issue should crop up in any organization that tries to create a library of standard components which implement the organization’s visual design language. Suppose your site’s designer has created a cool button class as an HTML custom element, and you have used that to create an Add to Cart button. You write some script so the button can show inside the button the existing number of items in a customer’s online shopping cart (to the right of the button label, say). You now want to package up the Add to Cart button so that it can be used as a component in its own right. For flexibility, you want the button’s text label to vary in places.

Well-defined subclassing semantics are essential for creating a UI component library with a good separation of concerns. If you look at the class hierarchy depicted for DateComboBox (above), you’ll get a sense of the degree to which it’s possible to portion out specific roles to a small constellation of classes, such that each class can focus on just doing one thing really well.

I’m hoping that it is possible to take advantage of such subclassing semantics in HTML custom elements — but it’s not proving to be particularly easy.

The puzzle

The puzzle is to come up with an architecture for custom element subclasses that meets the following design criteria:

  1. An instance of a subclass is a proper instance of its base class. All the normal JavaScript stuff should work: property/method access should go up the prototype chain, and a subclass instance should report that it is an “instanceof” the base class. By default, the HTML <element> syntax permits an “extends” attribute to identify a base class, but a purely script-based solution that sets up the class hierarchy correctly is equally valid.
  2. A subclass can put stuff into an insertion point defined by the base class. That is, the subclass can fill in a slot (or slots) defined by a base class. In turn, the subclass should be able to redefine such an insertion point so that the subclass itself can be subclassed.
  3. Unless overridden, all base class behavior should function properly in an instance of the subclass. E.g., if the base class wires up an event handler, then this works as expected for subclass instances too.
  4. Base class properties/methods can be overridden by the subclass. A subclass’ property/method implementation should be able to invoke the base class’ implementation by whatever language means are necessary. (CoffeeScript provides sugar for this; plain JavaScript developers have alternate ways of achieving the same result.)
  5. The base class can be any HTML custom element class; the base class author shouldn’t have to do special work a priori to enable this kind of subclassing. This ensures a Quetzal author can always use someone else’s element class as a base class — even if that other person has never heard of Quetzal.

A successful solution needs to meet all five of these criteria. So far, the approaches I’ve tried can satisfy at most four at a time.

Example

Let’s walk through a example from the small set of custom elements currently shown on the Quetzal home page. This set includes a base element class called quetzal-button that shows its content inside a button, and another element class called icon-button which adds an icon to the plain-button content. For clarity, here let’s just call that base class plain-button instead of quetzal-button, since the following source won’t actually involve Quetzal. In any event, we want markup like this:

<plain-button>Plain button</plain-button>
<icon-button icon="document.png">Icon button</icon-button>

… to produce something like this:

Buttons

Where icon-button is reusing all the styling and behavior from plain-button; it’s not duplicating the styling and behavior. The challenge is to create the icon-button element so it both inherits (in the class sense) from plain-button and extends the visual representation of plain-button.

A partial solution to filling in base class insertion points

The first challenge is: how can icon-button add elements to the content shown by plain-button? Some approaches:

  1. We could try to apply the template for both plain-button and icon-button to the same host element. The Shadow DOM spec supports multiple shadow trees attached to the same host. This feature alone is insufficient for the above example. Unless one shadow tree takes care to incorporate the other somehow, the most recently-added shadow subtree wins. If plain-button renders last, we get a button but no icon; if icon-button renders last, we get an icon but no button.
  2. We can include a <shadow> element in the template for icon-button, and ensure plain-button renders its shadow subtree first. The <shadow> element allows icon-button to effectively include the representation for plain-button. Unfortunately, this inclusion effectively wraps the base representation, rather than filling it in. An icon-button does get a button and an icon, but the icon and content render outside an empty button: shadow
  3. We can have icon-button create an instance of its own base class, then have that instance contain the icon and the icon-button’s own content. (This approach is based on a suggestion from Shadow DOM spec author Dimitri Glazkov.)

Approach #3 does what we want from a strictly visual perspective. (Behavior is a separate matter.) The source for a Polymer element version of this approach looks something like:

<element name="plain-button" extends="button">
    <template>
        <button>
            <content></content>
        </button>
    </template>
    …
</element>

<element name="icon-button" extends="plain-button" attributes="icon">
    <template>
        <plain-button>
            <img src="{{icon}}">
            <content></content>
        </plain-button>
    </template>
    …
</element>

See a live example of this. It should work in most browsers, but use Chrome if you want to see it working with real Shadow DOM. (Note: a Chrome bug prevents the buttons from responding correctly to mouse interactions. The buttons do receive mouse events, but for now it just doesn’t look that way.)

If you open the example and inspect it in Chrome dev tools, you’ll see that an icon-button has a shadow subtree containing a plain-button; the plain-button contains its own shadow subtree. Content of an <icon-button> element is therefore distributed twice: once into the <plain-button> element, and then again into the native <button> element. The ability of Shadow DOM to distribute nodes multiple times is called reprojection. (Or, at least, it meets the definition of reprojection as I understand it: “when an insertion point is a child node of another shadow host”.)

Inheritance versus containment

Unfortunately, while this approach looks right, it doesn’t behave quite right. An icon-button here is not only an instance of plain-button, it also contains a plain-button. That’s problematic.

  1. When icon-button instantiates its inner plain-button, the inner button has no way to know its relationship to icon-button. Among other things, this means icon-button can’t easily override behavior defined by plain-button (one of the design criteria above). In the JS Bin example, you can turn on the Console pane to see debug output. The plain-button element class defines a readyCallback (in Polymer, “ready”) that invokes a base class method called log(). The icon-button class overrides that log() method, but because the inner plain-button is just that — a plain-button — its readyCallback will invoke the base plain-button implementation of log() instead of icon-button’s specialized log() implementation. Running the demo invokes log() four times, when: 1) creating an instance of plain-button to use as the prototype for icon-button, 2) creating the visible plain-button with text “Plain button”, 3) creating the inner plain-button used by the visible icon-button, and 4) creating the visible icon-button with text “Icon button”. It’s #3 and #4 together that are the problem: what we really wanted to do is invoke icon-button’s log() implementation once.
  2. Automatic element references (a la Polymer, and also in Quetzal) aren’t inherited by default. If plain-button defines an element with id #foo, then plain-button methods can access that element via the automatic reference this.$.foo. Similarly, we want an icon-button to have access to the same reference this.$.foo defined by the base class. (Or, at least, we can debate whether such automatic references should be treated as “private” or “protected”, but it seems to me that “protected” would be useful.) It’s possible to work around this particular issue for a known set of frameworks — that is, Quetzal could workaround this problem for its own classes, and perhaps for those defined by Polymer — but it wouldn’t work in the general case of an unknown framework.
  3. It’s easy to end up in situations where both icon-button and plain-button are duplicating work. Suppose an icon-button method invokes a super-method of the same name defined by plain-button, and suppose the base implementation of that method performs expensive work or obtains a reference to some resource. When the inner plain-button is instantiated, it might do that work — and then the same work or allocation might be repeated by the outer icon-button when it invokes the super-method. Conventions could be established to avoid this, but it would complicate otherwise simple situations, and again make it hard to use subclass elements from other frameworks.
  4. As a common case of the above point, if the inner plain-button wires up an event handler, then it’s easy to end up in situations where the event is bound by both the inner plain-button and the outer icon-button. If the event bubbles up from something inside plain-button (a click on the button, say), you would end up handling the same event twice.

We could try to simplify things by just containing an plain-button, and not deriving from it. This forces us to give up one of the original design criteria outlined above: an instance of icon-button wouldn’t actually appear to be an “instanceof” plain-button. Moreover, if plain-button defined some attributes (“disabled”, say), icon-button would have to explicitly handle those too and forward their implementation to the inner plain-button.

We could have icon-button create a placeholder element (a <div>, say), create a shadow root for it, and populate that root with a copy of plain-button’s template but without actually instantiating that inner element as a real, live plain-button. This is the approach that Quetzal currently uses. It solves a number of problems, but is dependent on knowing how a given base element class works. Quetzal reaches into the base class’ implementation to obtain its template and then clones it, which might not be possible with other frameworks. This violates one of the design criteria above.

We could create a temporary instance of plain-button elsewhere, then clone just its contents into the icon-button instance. This avoids requiring detailed knowledge of what the base class is doing. But it could also result in subtle problems. E.g., the base class might not be expecting to have to serialize all its state into its shadow subtree, in which case the cloned content might not represent a coherent instance of the base class.

Looking for suggestions

This post is effectively a form of rubber duck debugging. The simple act of writing this up has forced me to better understand the problem, and led to consideration of alternate lines of attack. The puzzle remains unsolved, however. Given my understanding of custom elements, and the design criteria for the puzzle above, I’m not sure whether a solution exists.

It’s theoretically possible I’ve hit some limit in the expressiveness permitted to custom elements in their current state. Perhaps that limitation could be addressed. If not, I’d  have to write off a big chunk of the solutions used by the QuickUI Catalog and QuickUI apps, and find alternate ways of meeting the same needs.

I’m hoping, however, that I’m just missing something. If you have some passing familiarity with HTML custom elements and Shadow DOM, and have ideas about how to approach this problem within the existing technology, I’d love to hear them!

Advertisements

One thought on “Puzzle: Define HTML custom element subclasses that can fill in base class insertion points

  1. I thought I’d provide some supporting data to help make the case that the component feature described in this post — filling in base class insertion points — is highly useful. I’ve done a quick analysis of four component-based libraries and applications built with the QuickUI framework, two public and two proprietary. I was curious what percentage of components in a given project filled in a base class in the manner described in this post. A relatively high percentage of components taking advantage of this facility would suggest that the feature is useful, and worth accommodating in web components.

    Here are the percentage of base class-filling components found in four real applications (i.e., not tests, demos, etc.):
    12% (12 of 94 components ) in the QuickUI Catalog, a library of general purpose components.
    23% (7 of 30 components) in QuickUI.org, a web site for an open source project.
    33% (5 of 15 components) in a simple mobile site for a tech startup.
    41% (20 of 49 components) in a visually rich, browser-based interactive game.

    23% (44 of 188 components) is the average across the four projects. I.e., roughly a quarter of components in this sample set take advantage of this base class-filling feature in QuickUI.

    I should mentioned that these percentages are calculated by looking at *all* components in these apps. If one focuses on just those components which are inheriting from a class that defines visible elements, then the percentage of subclasses that want to fill in base class insertion points will probably much higher — likely over 50%. So, if you’re creating a component that extends a class that itself populates the (shadow) DOM with visible elements, odds are that subclass wants to fill in insertion points defined by the base class. In other words, the technique addressed in this post is likely the norm, not the exception.

    Stepping back and looking at these apps, I notice a rough correlation between the “specialness” of an app’s user experience and the degree to which it’s useful to use this base class-filling approach. The comparatively low percentage of the QuickUI Catalog stems from the fact that the Catalog is a library of general-purpose base classes. Many of those classes don’t inherit from another class (aside from the base Control class). QuickUI.org is a fairly plain web site, and makes modest use of this feature to specialize the appearance and/or behavior of a handful of base classes from the QuickUI Catalog. The tech startup’s mobile site is a bit more visually distinctive; even though it’s a smaller app, it has stronger branding. Finally, in the browser-based game, visual aesthetics and unique behavior make up a significant part of the app’s user experience — as opposed to pure functionality. So it’s not surprising that the game makes the heaviest use of this feature, which is fundamentally about making it easy to create specialized user interface elements.

    Of the four samples listed above, the game is probably the most representative of the demands of a mainstream consumer web app. If one were to look at component-based apps in production several years from now, one would probably find that the 23% average given above turned out to be a low estimate.

    Another way to consider the percentages above is to say that the figure represents the portion of app UI code which today *cannot* be easily ported from QuickUI to web components. This is the sort of situation one finds when one has to port an app whose code takes advantage of an underlying programming language feature (generics, say), and tries to port to a language without such a feature. It’s not that the port is impossible, but that the new language just can’t express something which the old language could.

Comments are closed.