Element Timing API

1. Introduction

This section is non-normative.

Knowing when critical elements are displayed on screen is key to understanding page load performance. While fast rendering of the essential components is not sufficient for a satisfactory loading experience, it is necessary. Therefore, monitoring these rendering timestamps is important to improve and prevent regressions in page loads.

This specification gives developers and analytics providers an API to measure rendering timestamps of critical elements. There is currently no good way to measure these timestamps for real users. Existing approaches would require either registering observers too early or significant DOM manipulation. These approaches are discussed on the § 4 Security & privacy considerations section.

Web developers are the experts in critical user interactions for their sites, so they should be allowed to tell the user agent which are the elements they care about. Thus, this API exposes rendering timing information about web-developer-annotated elements.

1.1. Elements exposed

The Element Timing API supports timing information about the following elements:

• img elements.

• image elements inside an svg.

• The poster images of video elements.

• Elements with a background-image.

• Groups of text nodes, which are aggregated as described in § 3.2 Modifications to the CSS specification.

Elements that have a "elementtiming" content attribute are reported in the report image element timing and the report text element timing algorithms.

1.2. Usage example

The following example shows an image that is registered for observation via its elementtiming attribute, and an observer gathering the timing information.

<img... elementtiming='foobar'/>
<p elementtiming='important-paragraph'>This is text I care about.</p>
...
<script>
const observer = new PerformanceObserver((list) => {
  let perfEntries = list.getEntries();
  // Process the entries by iterating over them.
});
observer.observe({type: 'element', buffered: true});
</script>

The following are sample elements whose rendering timestamps could be measured by using this API and which should be compared to page navigation:

• The images in the image carousel of a shopping site.

• The main photo in a story of a news site.

• The title of a blog post.

• The first paragraph in an entry of an encyclopedia site.

The API could have use cases outside of page load by comparing the rendering timestamps with input timestamps. For example, developers could monitor the time it takes for a widget to show up after a click that triggers it.

2. Element Timing

Element Timing involves the following new interfaces:

2.1. PerformanceElementTiming interface

[Exposed=Window]
interface PerformanceElementTiming : PerformanceEntry {
    readonly attribute DOMHighResTimeStamp renderTime;
    readonly attribute DOMHighResTimeStamp loadTime;
    readonly attribute DOMRectReadOnly intersectionRect;
    readonly attribute DOMString identifier;
    readonly attribute unsigned long naturalWidth;
    readonly attribute unsigned long naturalHeight;
    readonly attribute DOMString id;
    readonly attribute Element? element;
    readonly attribute DOMString url;
    [Default] object toJSON();
};

A PerformanceElementTiming object reports timing information about one associated element.

Each PerformanceElementTiming object has these associated concepts, all of which are initially set to null:

• A request containing the image request (if the entry is for image content).

• An element containing the associated Element.

The associated concepts and some attributes for PerformanceElementTiming are specified in the processing model in § 3.5 Report image Element Timing and § 3.6 Report text Element Timing.

The entryType attribute’s getter must return the DOMString "element".

The name attribute’s getter must return the value it was initialized to.

The startTime attribute’s getter must return the value of this’s renderTime if it is not 0, and the value of this’s loadTime otherwise.

The duration attribute’s getter must return 0.

The renderTime attribute must return the value it was initialized to.

The loadTime attribute’s getter must return the the value it was initialized to.

The intersectionRect attribute must return the value it was initialized to.

The identifier attribute’s getter must return the value it was initialized to.

The naturalWidth attribute must return the value it was initialized to.

The naturalHeight attribute must return the value it was initialized to.

The id attribute’s getter must return the value it was initialized to.

The element attribute’s getter must run the get an element algorithm this’s element and null as inputs.

Note: This means that an element that is no longer descendant of the Document will no longer be returned by element's attribute getter.

The url attribute’s getter must perform the following steps:

Note: The URL is trimmed for data URLs to avoid excessive memory in the entry.

3. Processing model

Note: A user agent implementing the Element Timing API would need to include "element" in supportedEntryTypes for Window contexts. This allows developers to detect support for element timing.

3.1. Modifications to the DOM specification

This section will be removed once the [DOM] specification has been modified.

We extend the Element interface as follows:

partial interface Element {
    [CEReactions] attribute DOMString elementTiming;
};

The elementTiming attribute must reflect the element’s "elementtiming" content attribute.

Each Element has an associated image request which is initially null. When the processing model for an Element element of type HTMLImageElement, SVGImageElement, or HTMLVideoElement creates a new image resource (e.g., to be displayed as an image or poster image), element’s associated image request is set to the image request of the created image resource.

Note: Every image resource that is obtained from a URL whose scheme is equal to "data" has an associated image request which is not fetched but still needs to be loaded. This request can be the associated image request of an Element.

Note: The current language is vague since it does not point to specific algorithms. This can be made more rigorous when the corresponding processing models have a more unified processing model.

Each Element has a set of owned text nodes, which is initially an empty ordered set.

Each Document has images pending rendering, a list of triples (Element, image request, DOMHighResTimeStamp) which are considered loaded but not yet rendered. When an Element element’s associated image request has become completely available, run the algorithm to process an image that finished loading passing in element and its associated image request as inputs.

Each Document also has a set of elements with rendered text, which is initially an empty, ordered set.

3.2. Modifications to the CSS specification

When the user agent is executing the painting order, it must populate the set of owned text nodes of the painted Elements so that the following is true:

NOTE: A user agent might want to use a stack to efficiently compute the set of owned text nodes while implementing the painting order.

Every Element has a list of associated background image requests which is initially an empty array. When the processing model for the Element element’s style requires a new image resource (to be displayed as background image), the image request created by the new resource is appended to element’s associated background image requests. Whenever an image request in an Element element’s associated background image requests becomes completely available, run the algorithm to process an image that finished loading with element and image request as inputs.

NOTE: we assume that there is one image request for each Element that a background-image property affects and for each URL that the background-image property specifies. So, for example, if there is a style with two URLs affecting all divs, and there are two divs, then there will be four image requests. This means that a single background-image property could produce multiple PerformanceElementTiming entries because it can affect multiple elements and because it can specify multiple URLs.

3.3. Modifications to the HTML specification

This section will be removed once the [HTML] specification has been modified.

In the update the rendering step of the event loop processing model, add the following substep at the end:

1. For each fully active Document in docs, run the element timing processing algorithm passing in the Document and now.

3.4. Process image that finished loading

3.5. Report image Element Timing

3.6. Report text Element Timing

3.7. Element Timing processing

3.8. Get an element algorithm

4. Security & privacy considerations

This API exposes some information about cross-origin images. In particular, images that do not pass the timing allow check still have their resource load time exposed, which could be a source of privacy concerns.

However, this is considered to not add new attacks to the web platform because the ResourceTiming API exposes a similar timestamp already. In addition, the onload handler exposes load timing when it is available, and the resource load time is a close proxy to this. The current high resolution time computed at the beginning of the onload handler would provide the image load time. We choose to expose the loadTime because it is very easy to obtain even without an onload handler. In addition, we believe any fix to remove the leak provided by image onload handlers or ResourceTiming could also fix the leak provided by this API.

The renderTime (display timestamp) can also be polyfilled via the PaintTiming API. To do this, add an iframe that contains trivial content on the onload handler of the target image or text content. Then, query the first paint of that iframe to obtain the rendering timestamp of the content. This is quite inefficient and the polyfill itself might affect the timing obtained. Due to the difficulty in obtaining this information today, we choose not to expose the display timestamp for images that fail the timing allow check. For clarity, here is a code snippet using the PaintTiming API:

// In the attacker frame.
<iframe src=attack.html></iframe>
<script>
    window.onmessage = e => {
        let timestamp = e.data;
        // Acquired the display timestamp for 'victim.jpg'!
    }
</script>

// In the attack.html iframe.
<img src='victim.jpg'/>
<script>
    // Wait until onload or some time when the PaintTiming entries will be visible.
    onload() => {
        let entry = performance.getEntriesByType('paint')[0];
        top.postMessage(entry.startTime, '*');
    }
</script>

The other nontrivial parameter being exposed here is the intersectionRect. This can already be polyfilled, for example using IntersectionObserver. The polyfill process would be similar: add an IntersectionObserver on the onload handler of the target image or text content. This solution is inefficient because it requires registering the observer once the content has loaded, but it should still provide the same level of accuracy. For images, we compute the intersectionRect once the image has loaded if it does not pass the timing allow check. Computing it at this point in time allows exposing the entry at that time. If we were to compute the rect only until the image is fully displayed, we’d only be able to expose the entry after that time.

If we do not want to expose the rendering timetamp of an image, it’s preferable to dispatch the entry to the PerformanceObserver right away. Suppose we waited and exposed all the entries during the element timing processing algorithm. An attacker could infer nontrivial information about the rendering timestamp of an image. It would do so by only observing the timing for that image. Even though the timestamp is not exposed as a member of the PerformanceElementTiming entry received, the fact that we wait until the next update the rendering step means that the attacker can distinguish between a very slow rendering time and a very fast rendering time by measuring the time at which it received the entry. This would unintentionally leak some of the display timing of the image.