Infrequently Noted

Alex Russell on browsers, standards, and the process of progress.

Reactive Data With Modern JavaScript

EventTarget and Proxies FTW

A silly little PWA has been brewing over the past couple of weekends to make a desktop-compatible version of a mobile-only native app using Web Bluetooth.

I'm incredibly biased of course, but the Project Fugu 🐡 APIs are a lot of fun. There's so much neat stuff we can build in the browser now that HID, Serial, NFC, Bluetooth and all the rest are available. It has been a relaxing pandemic distraction to make time to put them through their paces, even if developing them is technically the day job too.

Needless to say, browsers that support Web Bluetooth are ultra-modern[1]. There's no point in supporting legacy browsers that don't have this capability, which means getting to use all the shiny new stuff; no polyfills or long toolchains. Fun!

In building UI with lit-html, trigger rendering without littering code with calls to render(...) can be a challenge. Lots of folks use data store libraries that provide a callback life-cycle, but they seem verbose and I only want things to be as complicated as they are complex.

Being reactive to data changes without a passel of callbacks only needs:

That's it! Before modern runtimes, we needed verbose, explicit API surface. But it's 2021, and we can do more with less now thanks to Proxies and subclassable EventTarget.

Hewing to the "data down, events up" convention of Web Components, here's a little function my small app is using instead of a "state management" library:

* proxyFor.js
* Utilities for wrapping an object or graph in a Proxy
* that dispatches `ondatachange` events to an EventTarget.

// Bookeeping
let objToProxyMap = new WeakMap();

const doNotProxy = Symbol.for("doNotProxy");
const isProxied = Symbol.for("isProxied");

let inDoNotProxy = (obj, prop) => {
if (prop === doNotProxy) { return true; }
if (obj) {
if (prop && obj[doNotProxy]) {
if (obj[doNotProxy].indexOf(prop) !== -1) {
return true;
return false;

let shouldNotProxy = (thing) => {
return (
(thing === null) ||
(typeof thing !== "object") ||

// proxyFor() supports two modes, "currentOnly" or "shallow"
// proxying (which is the default) avoids creating new wrappers
// for objects extracted from properties on the passed object.
// This is faster but requires that developers understand the
// limitations.
// The other mode, deep proxies, create new wrappers around
// each object returned from any property. This mode recursively
// wraps sub-objects in proxies that notify the passed
// EventTarget.
// Note that in the case of an object that is itself an
// EventTarget, one can pass it as both the first and second
// argument to notify on it directly. This also provides
// flexibility so that a notifying proxy can be wired up
// to send change events to a *different* object.
export let proxyFor = (thing, eventTarget=null,
currentOnly=true, path=[]
) => {

// Bail if not an object, or if already proxied
if (shouldNotProxy(thing)) {
return thing;
if (!eventTarget) {
"Missing eventTarget. Could not proxy for", thing);
return thing;

let dataProperties = currentOnly ?
new Set(Object.keys(thing)) : null;

let inDataProperties = (prop) => {
return (!currentOnly || dataProperties.has(prop));

let p = new Proxy(thing, {

get: function(obj, prop, receiverProxy) {
// Avoid any potential re-wrapping
if (prop && prop === isProxied) { return true; }

let value = Reflect.get(...arguments);
let valueType = (typeof value);

if (valueType === "undefined") { return; }

if (valueType === "object" &&
value !== null &&
value[isProxied]) {
return value;

if (objToProxyMap.has(value)) {
return objToProxyMap.get(value);

// Avoid `this` confusion for functions
if ((valueType === "function") &&
(prop in EventTarget.prototype)) {
return value.bind(obj);

// Do not create wrappers for sub-objects
// in these cases
if (inDoNotProxy(obj, prop) ||
inDataProperties(prop)) {
return value;

// Handle object trees by returning a wrapper
return proxyFor(value, eventTarget, currentOnly,

set: function(obj, prop, value) {
if (!dataProperties || dataProperties.has(prop)) {
let evt = new CustomEvent("datachange",
{ bubbles: true, cancelable: true, }
evt.oldValue = thing[prop];
evt.value = value;
evt.dataPath = path.concat(prop); = prop;

obj[prop] = value;

return true;

objToProxyMap.set(thing, p);

return p;

// For deep objects that are themselves EventTargets
export let deepProxyFor = (thing) => {
return proxyFor(thing, thing, false);

One way to use this is to mix it in with a root object that is itself an EventTarget:

// AppObject.js

import { proxyFor, deepProxyFor } from "./proxyFor.js";

// In modern runtimes, EventTarget is subclassable
class DataObject extends EventTarget {

aNumber = 0.0;
aString = "";
anArray = [];
// ...

constructor() {
// Cheeky and slow, but works
return deepProxyFor(this);

export class AppObject extends DataObject {
// We can handle inherited properties mixed in
counter = 0;

doStuff() {
this.aNumber += 1.1;
this.aString = (this.aNumber).toFixed(1) + "";
this.anArray.length += 1; // Handled

The app creates instances of AppObject and subscribes to datachange events to drive UI updates once per frame:

<script type="module">
import { html, render } from "lit-html";
import { AppObject } from "./AppObject.js";

let app = new AppObject();

let mainTemplate = (obj) => {
return html`
A Number:
A String: "

<button @click=
${() => { obj.counter++; }}>+</button>
<button @click=
${() => { obj.counter--; }}>-</button>

// Debounce to once per rAF
let updateUI = (() => {
let uiUpdateId;
return function (obj, tmpl, node, evt) {
if (!node) { return; }
if (uiUpdateId) {
uiUpdateId = null;
uiUpdateId = requestAnimationFrame(() => {
// Logs/renders once per frame
render(tmpl(obj), node);

// Wire the template to be re-rendered from data
let main = document.querySelector("main");

app.addEventListener("datachange", (evt) => {
// Not debounced, called in quick succession by
// setters in `doStuff`
updateUI(app, mainTemplate, main, evt);

setInterval(app.doStuff.bind(app), 1000);
<!-- ... -->

This implementation debounces rendering to once per requestAnimationFrame(), which can be extended/modified however one likes.

There are other caveats to this approach, some of which veritably leap off the page:

In general, we win idiomatic object syntax for most operations at the expense of breaking private properties, but for a toy, it's a fun start.

Here's another way to use this, routing updates on a data object through an existing element:

<!-- A DOM-driven reactive incrementer -->
<script type="module">
import { proxyFor } from "./proxyFor.js";

let byId = window.byId =

let count = byId("count"); = proxyFor({ value: 0 }, count);

// Other parts of the document can listen for this
count.addEventListener("datachange", (evt) => {
count.innerText = evt.value;

Current count:
<span id="count">0</span>
<button id="increment"

<button id="decrement"


You can try a silly little test page, wonder at the obligatory reactive incrementer demo, or the Svelte Store API implemented using proxyFor.

  1. Sorry iOS users, there are no modern browsers available for your platform because Apple is against meaningful browser-choice, no matter what colour lipstick vendors are allowed to put on WebKit. ↩ī¸Ž

The Mobile Performance Inequality Gap, 2021

An update on phones, networks, browsers, and the new baseline scenario for web performance.

TL;DR: A lot has changed since 2017 when we last estimated a global baseline resource budget of 130-170KiB per-page. Thanks to progress in networks and browsers (but not devices), a more generous global budget cap has emerged for sites constructed the "modern" way. We can now afford ~100KiB of HTML/CSS/fonts and ~300-350KiB of JS (gzipped). This rule-of-thumb limit should hold for at least a year or two. As always, the devil's in the footnotes, but the top-line is unchanged: when we construct the digital world to the limits of the best devices, we build a less usable one for 80+% of the world's users.

Way back in 2016, I tried to raise the alarm about the causes and effects of terrible performance for most users of sites built with popular frontend tools. The negative effects were particularly pronounced in the fastest-growing device segment: low-end to mid-range Android phones.

Bad individual experiences can colour expectations of the entire ecosystem. Your company's poor site performance can manifest as lower engagement, higher bounce rates, or a reduction in conversions. While this local story is important, it isn't the whole picture. If a large enough proportion of sites behave poorly, performance hysteresis may colour user views of all web experiences.

Unless a site is launched from the home screen as a PWA, sites are co-mingled. Pages are experienced as a series of taps, flowing effortlessly across sites; a river of links. A bad experience in the flow is a bad experience of the flow, with constituent parts blending together.[1]

If tapping links tends to feel bad...why keep tapping? It's not as though slow websites are the only way to access information. Plenty of native apps are happy to aggregate content and serve it up in a reliably fast package, given half a chance. The consistency of those walled gardens is a large part of what the mobile web is up against — and losing to.

The Mobile Web: MIA

Poor performance of sites that link to and from yours negatively impacts engagement on your site, even if it is consistently snappy. Live by the link, die by the link.

The harmful business impact of poor performance is constantly re-validated. Big decreases in performance predictably lead to (somewhat lower) decreases in user engagement and conversion. The scale of the effect can be deeply situational or hard to suss out without solid metrics, but it's there.

Variance contributes another layer of concern; high variability in responsiveness may create effects that perceptually dominate averages, or even medians. If 9 taps in 10 respond in 100ms, but every 10th takes a full second, what happens to user confidence and engagement? These deep-wetware effects and their cross-origin implications mean that your site's success is, partially, a function of the health of the commons.

From this perspective, it's helpful to consider what it might take to set baselines that can help ensure minimum quality across link taps, so in 2017 I followed up with a post sketching a rubric for thinking about a global baseline.

The punchline:

The default global baseline is a ~$200 Android device on a 400Kbps link with a 400ms round-trip-time ("RTT"). This translates into a budget of ~130-170KB of critical-path resources, depending on composition — the more JS you include, the smaller the bundle must be.

A $200USD device at the time featured 4-8 (slow, in-order, low-cache) cores, ~2GiB of RAM, and pokey MLC NAND flash storage. The Moto G4, for example.

The 2017 baseline represented a conservative, but evidence-driven, interpretation of the best information I could get regarding network performance, along with trend lines regarding device shipment volumes and price points.[2] Getting accurate global information that isn't artificially reduced to averages remains an ongoing challenge. Performance work often focuses on high percentile users (the slowest), after all.

Since then, the metrics conversation has moved forward significantly, culminating in Core Web Vitals, reported via the Chrome User Experience Report to reflect the real-world experiences of users.

Devices and networks have evolved too:

An update on mobile CPUs and the Performance Inequality Gap:

Mid-tier Android devices (~$300) now get the single-core performance of a 2014 iPhone and the multi-core perf of a 2015 iPhone.

The cheapest (high volume) Androids perform like 2012/2013 iPhones, respectively.

Image from Tweet

Meanwhile, developer behaviour offers little hope:

Median mobile JavaScript payloads have only grown since 2016, now hovering above 400KiB of script transferred, or nearly 2.5MiB of uncompressed JS.
Median mobile JavaScript payloads have only grown since 2016, now hovering above 400KiB of script transferred, or nearly 2.5MiB of uncompressed JS.

A silver lining on this dark cloud is that mobile JavaScript payload growth paused in 2020. How soon will low-end and middle-tier phones be able to handle such chonky payloads? If the median site continued to send 3x the recommended amount of script, when would the web start to feel usable on most of the world's devices?

Here begins our 2021 adventure.

Hard Reset

To update our global baseline from 2017, we want to update our priors on a few dimensions:

Content Is Dead, Long Live Content

Good news has consistently come from the steady pace of browser progress. The single largest improvements visible in traces come from improved parsing and off-thread compilation of JavaScript. This step-change, along with improvements to streaming compilation, has helped to ensure that users are less likely to notice the unreasonably-sized JS payloads that "modern" toolchains generate more often than not. Better use of more cores (moving compilation off-thread) has given sites that provide HTML and CSS content a fighting chance of remaining responsive, even when saddled with staggering JS burdens.

A steady drumbeat of improvements have contributed to reduced runtime costs, too, though I fear they most often create an induced demand effect.

The residual main-thread compilation, allocation, and script-driven DOM/Layout tasks pose a challenge for delivering a good user experinece.[3] As we'll see below, CPUs are not improving fast enough to cope with frontend engineers' rosy resource assumptions. If there is unambiguously good news on the tooling front, multiple popular tools now include options to prevent sending first-party JS in the first place (Next.js, Gatsby), though the JS community remains in stubborn denial about the costs of client-side script. Hopefully, toolchain progress of this sort can provide a more accessible bridge as we transition costs to a reduced-script-emissions world.

4G Is A Miracle, 5G Is A Mirage

The 2017 baseline post included a small model for thinking about how to think about how various factors of a page's construction influence the likelihood of hitting a 5-second first load goal.

The hard floor of that model (~1.6s) came from the contributions DNS, TCP/IP, and TLS connection setup over a then-reasonable 3G network baseline, leaving only 3400ms to work with, fighting Nagle and weak CPUs the whole way. Adding just one extra connection to a CDN for a critical path resource could sink the entire enterprise. Talk about a hard target.

Four years later, has anything changed? I'm happy to report that it has. Not as much as we'd like, of course, but the worldwide baseline has changed enormously. How? Why?

Reliance Jio, mostly.

India has been the epicentre of smartphone growth in recent years, owing to the sheer size of its market and an accelerating shift away from feature phones which made up the majority of Indian mobile devices until as late as 2019. Continued projections of double-digit market growth for smartphones, on top of a doubling of shipments in the past 5 years, paint a vivid picture.

Key to this growth is the effect of Reliance Jio's entry into the carrier market. Before Jio's disruptive pricing and aggressive rollout, data services in India were among the most expensive in the world relative to income and heavily reliant on 3G outside wealthier "tier 1" cities. Worse, 3G service often performed like 2G in other markets, thanks to under-provisioning and throttling by incumbent carriers.

In 2016, Jio swept over the subcontinent like a monsoon dropping a torrent of 4G infrastructure and free data rather than rain.

Competing carriers responded instantly, dropping prices aggressively, leading to reductions in per-packet prices approaching 95%.

Jio's blanket 4G rollout blitzkrieg shocked incumbents with a superior product at an unheard-of price, forcing the entire market to improve data rates and coverage. India became a 4G-centric market sometime in 2018.

If there's a bright spot in our construction of a 2021 baseline for performance, this is it. We can finally upgrade our assumptions about the network to assume slow-ish 4G almost everywhere (pdf).

5G looks set to continue a bumpy rollout for the next half-decade. Carriers make different frequency band choices in different geographies, and 5G performance is heavily sensitive to mast density, which will add confusion for years to come. Suffice to say, 5G isn't here yet, even if wealthy users in a few geographies come to think of it as "normal" far ahead of worldwide deployment[4].

Hardware Past As Performance Prologue

Whatever progress runtimes and networks have made in the past half-decade, browsers are stubbornly situated in the devices carried by real-world users, and the single most important thing to understand about the landscape of devices your sites will run on is that they are not new phones.

This makes some intuitive sense: smartphones are not in their first year (and haven't been for more than a dozen years), and most users do not replace their devices every year. Most smartphone sales today are replacements (that is, to users who have previously owned a smartphone), and the longevity of devices continues to rise.

The worldwide device replacement average is now 33 months. In markets near smartphone saturation, that means we can expect the median device to be nearly 18 months old. Newer devices continue to be faster for the same dollar spent. Assuming average selling prices (ASPs) remaining in a narrow year-on-year band in most geographies[5], a good way to think of the "average phone" as being the average device sold 18 months ago. ASPs, however, have started to rise, making my prognostications from 2017 only technically correct:

The true median device from 2016 sold at about ~$200 unlocked. This year's median device is even cheaper, but their performance is roughly equivalent. Expect continued performance stasis at the median for the next few years. This is part of the reason I suggested the Moto G4 last year and recommend it or the Moto G5 Plus this year.

Median devices continue to be different from averages, but we can squint a little as we're abstracting over multi-year cohorts. The worldwide ASP 18 months ago was ~$300USD, so the average performance in the deployed fleet can be represented by a $300 device from mid-2019. The Moto G7 very much looks the part.

Compared to devices wealthy developers carry, the performance is night and (blinding) day. However shocking a 6x difference in single-thread CPU performance might be, it's nothing compared to where we should be setting the global baseline: the P75+ device. Using our little mental model for device age and replacement, we can naively estimate what the 75th percentile (or higher) device could be in terms of device price + age, either by tracking devices at half the ASP at half the replacement age, or by looking at ASP-priced devices 3/4 of the way through the replacement cycle. Today, either method returns a similar answer.

This is inexact for dozens of reasons, not least of all markets with lower ASPs not yet achieving smartphone saturation. Using a global ASP as a benchmark can further mislead thanks to the distorting effect of ultra-high-end prices rising while shipment volumes stagnate. It's hard to know which way these effects cut when combined, so we're going to make a further guess: we'll take half the average price and note how wrong this likely is[6].

So what did $150USD fetch in 2019?

Say hello to the Moto E6! This 2GiB RAM, Android 9 stalwart features the all-too classic lines of a Quad-core A53 (1.4GHz, small mercies) CPU, tastefully presented in a charming 5.5" package.

The smooth, dulcet tones of 2019's Moto E6.
The smooth, dulcet tones of 2019's Moto E6.

If those specs sound eerily familiar, it's perhaps because they're identical to 2016's $200USD Moto G4, all the way down to the 2011-vintage 28nm SoC process node used to fab the chip's anemic, 2012-vintage A53 cores. There are differences, of course, but not where it counts.

You might recall the Moto G4 as a baseline recommendation from 2016 for forward-looking performance work. It's also the model we sent several dozen of to Pat Meenan — devices that power to this day. There was no way to know that the already-ageing in-order, near-zero-cache A53 + too-hot-to-frequency-scale 28nm process duo would continue to haunt us five year on. Today's P75 devices tell the story of the yawning Performance Inequality Gap: performance for those at the top end continues to accelerate away from the have-nots who are perpetually stuck with 2014's hardware.

The good news is that chip progress has begun to move, if glacially, in the past couple of years. 28nm chips are being supplanted at the sub-$200USD price-point by 14nm and even 11nm parts[7]. Specs are finally improving quickly at the bottom of the market now, with new ~$135USD devices sporting CPUs that should give last year's mid-range a run for its money.

Mind The Gap

Regardless, the overall story for hardware progress remains grim, particularly when we recall how long device replacement cycles are:

<em>Tap for a larger version.</em><br>Updated Geekbench 4 single-core scores for each mobile price-point.
Tap for a larger version.
Updated Geekbench 4 single-core scores for each mobile price-point.

Recall that single-core performance most directly translates into speed on the web. Android-ecosystem SoC performance is, in a word, disastrous.

How bad is it?

We can think about each category in terms of years behind contemporary iPhone releases:

You're reading that right: single core Android performance at the low end is both shockingly bad and dispiritingly stagnant.

<em>Tap for a larger version.</em><br>Android ecosystem SoC's fare slightly better on multi-core performance, but the Performance Inequality Gap is growing there, too.
Tap for a larger version.
Android ecosystem SoC's fare slightly better on multi-core performance, but the Performance Inequality Gap is growing there, too.

The multi-core performance shows the same basic story: iOS's high-end and the most expensive Androids are pulling away from the volume-shipment pack. The fastest Androids predictably remain 18-24 months behind, owing to cheapskate choices about cache sizing by Qualcomm, Samsung Semiconductor, MediaTek, and the rest. Don't pay a lot for an Android-shaped muffler.

Chip design choices and silicon economics are the defining feature of the still-growing Performance Inequality Gap.

Things continue to get better and better for the wealthy, leaving the rest behind. When we construct a digital world to the limits of the best devices, the worse an experience we build, on average, for those who cannot afford iPhones or $800 Samsung flagships.

It is perhaps predictable that, instead of presenting a bulwark against stratification, technology outcomes have tracked society's growing inequality. A yawning chasm of disparities is playing out in our phones at the same time it has come to shape our economic and political lives. It's hard to escape thinking they're connected.

Developers, particularly in Silicon Valley firms, are definitionally wealthy and enfranchised by world-historical standards. Like upper classes of yore, comfort ("DX") comes with courtiers happy to declare how important comfort must surely be. It's bunk, or at least most of it is.

As frontenders, our task is to make services that work well for all, not just the wealthy. If improvements in our tools or our comfort actually deliver improvements in that direction, so much the better. But we must never forget that measurable improvement for users is the yardstick.

Instead of measurement, we seem to suffer a proliferation of postulates about how each new increment of comfort must surely result in better user experiences. But the postulates are not tied to robust evidence. Instead, experience teaches that it's the process of taking care to attend to those least-well-off that changes outcomes. Trickle-down user experience from developer-experience is, in 2021, as fully falsified as the Laffer Curve. There's no durable substitute for compassion.

A 2021 Global Baseline

It's in that spirit that I find it important to build to a strawperson baseline device and network target. Such a benchmark serves as a stand-in until one gets situated data about a site or service's users, from whom we can derive informed conclusions about user needs and appropriate resource budgets.

A global baseline is doubly important for generic frameworks and libraries, as they will be included in projects targeted at global-baseline users, whether or not developers have that in mind. Tools that cannot fit comfortably within, say, 10% of our baseline budget should be labelled as desktop-only or replaced in our toolchains. Many popular tools over the past 5 years have been blithely punctured these limits relative to the 2017 baseline, and the results have been predictably poor.

Given all the data we've slogged through, we can assemble a sketch of the browsers, devices, and networks today's P50 and P75 users will access sites from. Keep in mind that this baseline might be too optimistic, depending on your target market. Trust but verify.

OpenSignal's global report on connection speeds (pdf) suggest that WebPageTest's default 4G configuration (9Mbps w/ 170ms RTT) is a reasonable stand-in for the P75 network link. WPT's LTE configuration (12Mbps w/ 70ms RTT) may actually be too conservative for a P50 estimate today; something closer to 25Mbps seems to approach the worldwide median today. Network progress has been astonishing, particularly channel capacity (bandwidth). Sadly, data on latency is harder to get, even from Google's perch, so progress there is somewhat more difficult to judge.

As for devices, the P75 situation remains dire and basically unchanged from the baseline I suggested more than five years ago. Slow A53 cores — a design first introduced nearly a decade ago — continue to dominate the landscape, differing only slightly in frequency scaling (through improved process nodes) and GPU pairing. The good news is that this will change rapidly in the next few years. The median new device is already benefiting; modern designs and fab processes are powering a ramp-up in performance for a $300USD device, particularly in multi-core workloads. But the hardware future is not evenly distributed, and web workloads aren't heavily parallel.

Astonishingly, I believe this means that for at least the next year we should consider the venerable Moto G4 to still be our baseline. If you can't buy one new, the 2020-vintage Moto E6 does a stirring rendition of the classic; it's basically the same phone, after all. The Moto E7 Plus is a preview of better days to come, and they can't arrive soon enough.

"All models are wrong, but some are useful."

Putting it all together, where does that leave us? Back-of-the napkin, and accepting our previous targets of 5 seconds for first load and two seconds for second load, what can we afford?

Plugging the new P75 numbers in, the impact of network improvements are dramatic. Initial connection setup times are cut in half, dropping from 1600ms to 700ms, which frees up significant headroom to transmit more data in the same time window. The channel capacity increase to 9Mbps is enough to transmit more than 4MiB (megabytes) of content over a single connection in our 5 second window (1 Byte == 8 bits, so 9Mbps is just over ~1MBps). Of course, if most of this content is JavaScript, it must be compiled and run, shrinking our content window back down significantly. Using a single connection (thanks to the magic of HTTP/2), a site composed primarily of JavaScript loaded perfectly can, in 2021, weigh nearly 600KiB and still hit the first-load bar.

That's a very fine point to balance on, though. A single additional TCP/TLS connection setup in the critical path reduces the amount by 100KiB (as do subsequent critical-path network handshakes). Serialized requests for data, high TTFBs, and ever-present font serving issues make these upper limits far higher than what a site shooting for consistently good load times should be aiming for. In practice, you can't actually afford 600KiB of content if your application is build in the increasingly popular "single page app" style.

As networks have improved, client-side CPU time now dominates script download, adding further variability. Not all script parses and runs in the same amount of time for the same size. The contribution of connection setup also looms large.

If you want to play with the factors and see how they contribute, you can try this slightly updated version of 2017's calculator (launch a larger version in a new tab here):

<em>Tap to try the interactive version.</em>
Tap to try the interactive version.

This model of browser behaviour, network impacts, and device processing time is, of course, wrong for your site. It could be wrong in ways big and small. Perhaps your content is purely HTML, CSS, and images which would allow for a much higher budget. Or perhaps your JS isn't visually critical path, but still creates long delays because of delayed fetching. And nobody, realistically, can predict how much main-thread work a given amount of JS will cause. So it's an educated guess with some padding built in to account for worst-case content-construction (which is more common than you or I would like to believe).

Conservatively then, assuming at least 2 connections need to be set up (burning ~1400 of our 5000ms), and that script resources are in the critical path, the new global baseline leaves space for ~100KiB (gzipped) of HTML/CSS/fonts and 300-350KiB of JavaScript on the wire (compressed). Critical path images for LCP, if any, need to subtract their transfer size from one of these buckets, with trades against JS bundle size potentially increasing total download budget, but the details matter a great deal. For "modern" pages, half a megabyte is a decent hard budget.

Is that a lot? Well, 2-4x our previous budget, and as low-end CPUs finally begin to get faster over the next few years, we can expect it to loosen further for well-served content.

Sadly, most sites aren't perfectly served or structured, and most mobile web pages send more than this amount of JS. But there's light at the end of the tunnel: if we can just hold back the growth of JS payloads for another year or two — or reverse the trend slightly — we might achieve a usable web for the majority of the world's users by the middle of the decade.

Getting there involves no small amount of class traitorship; the frontend community will need to value the commons over personal comfort for a little while longer to ease our ecosystem back toward health. The past 6 years of consulting with partner teams has felt like a dark remake of Groundhog Day, with a constant parade of sites failing from the get-go thanks to Framework + Bundler + SPA architectures that are mismatched to the tasks and markets at hand. Time (and chips) can heal these wounds if we only let it. We only need to hold the line on script bloat for a few years for devices and networks to overtake the extreme, unconscionable excesses of the 2010's.

Baselines And Budgets Remain Vital

I mentioned near the start of this too-long piece that metrics better than Time-to-Interactive have been developed since 2017. The professionalism with which the new Core Web Vitals (CWV) have been developed, tested, and iterated on inspires real confidence. CWV's constituent metrics and limits work to capture important aspects of page experiences as users perceive them on real devices and networks.

As RUM (field) metrics rather than bench tests, they represent ground truth for a site, and are reported in aggregate by the Chrome User Experience Report (CrUX) pipeline. Getting user-population level data about the performance of sites without custom tooling is a 2016-era dream come true. Best of all, the metrics and limits can continue to evolve in the future as we gain confidence that we can accurately measure other meaningful components of a user's journey.

These field metrics, however valuable, don't yet give us enough information to guide global first-best-guess making when developing new sites. Continuing to set performance budgets is necessary for teams in development, not least of all because conformance can be automated in CI/CD flows. CrUX data collection and first-party RUM analytics of these metrics require live traffic, meaning results can be predicted but only verified once deployed.

Both budgets and RUM analytics will continue to be powerful tools in piercing the high-performance privilege bubble that surrounds today's frontend discourse. Grounding our choices against a baseline and measuring how it's going for real users are the only proven approaches I know that reliably help teams take the first step toward frontend's first and highest calling: a web that's truly for everyone.

  1. One subtle effect of browser attempts to avoid showing users blank white screens is that optimisations meant to hide latency may lead users to misattribute which sites are slow.

    Consider a user tapping a link away from your site to a slow, cross-origin site. If the server takes a long time to respond, the referring page will be displayed for as long as it takes for the destination server to return with minimally parseable HTML content. Similarly, if a user navigates to your site, but the previous site's onunload handlers take a long time to execute (an all too common issue), no matter how fast your site responds, the poor performance of the referring page will make your site appear to load slowly.

    Even without hysteresis effects, the performance of referrers and destinations can hurt your brand. ↩ī¸Ž

  2. A constant challenge is the tech world's blindness to devices carried by most users. Tech press coverage is over-fitted to the high-end phones sent for review (to say nothing of the Apple marketing juggernaut), which doesn't track with shipment volumes.

    Smartphone markers (who are often advertisers) want attention for the high-end segment because it's where the profits are, and the tech press rarely pushes back. Column inches covering the high-end remains heavily out of proportion to sales by price-point. In a world with better balance, most articles would be dedicated to announcements of mid-range devices from LG and Samsung.

    International tech coverage would reflect that, while Apple finally pierced 50% US share for the first time in a decade at the end of 2020, iOS smartphone shipments consistently max out at ~20% worldwide, toiling near 15% most quarters.

    There's infinitely more to say about these biases and how they manifest. It's constantly grating that Qualcomm, Samsung Semiconductor, MediaTek, and other Android SoC vendors get a pass (even at the high end) on their terrible price/performance.[7:1]

    Were we to suffer an outbreak of tenacious journalism in our tech press, coverage of the devastatingly bad CPUs or the galling OS update rates amongst Android OEMs might be among the early signs. ↩ī¸Ž

  3. In the worst cases, off-thread compilation has no positive impact on client-side rendered SPAs. If an experience is blocked on a script to begin requesting data or generating markup from it, additional cores can't help. Some pages can be repaired with a sprinkling of <link rel=preload> directives.

    More frequently, "app"-centric architectures promulgated by popular framework starter kits leave teams with immense piles of JS to whittle down. This is expensive, gruelling work in the Webpack mines; not all the "developer experience" folks signed up for.

    The React community, in particular, has been poorly served by a lack of guidance, tooling, and support from Facebook. Given the scale of the disaster that has unfolded there, it's shocking that in 2021 (and after recent re-development of the documentation site) there remains no guidance about bundle sizes or performance budgets in React's performance documentation.

    Facebook, meanwhile, continues to staff a performance team, sets metrics and benchmarks, and generally exhibits the sort of discipline about bundle sizes and site performance that allows teams to succeed with any stack. Enforced latency and performance budgets tend to have that effect.

    FB, in short, knows better. ↩ī¸Ž

  4. As ever, take the money and introductions VCs offer, but interpret their musings about what "everyone knows" and "how things are" like you might Henry Kissinger's geopolitical advise: toxic, troubling — and above all — unfailingly self-serving. ↩ī¸Ž

  5. Device Average Selling Price (APSs) are both directionally helpful and deeply misleading.

    Global ASPs suffer all the usual issues of averages. A slight increase or decrease in the global ASP can mask extreme variance between geographies, major disparities in volumes, and shifts in prices at different deciles.

    We see this misfeature in the wild through press treatment of large headline increases in ASP and iOS share when Android device sales fall even slightly. Instead of digging into the shape of the distribution and relative time series inputs, the press regularly writes this up as the sudden success of some new Apple feature. That's always possible, but a slightly stylised version of these facts more easily pass Occam's test: wealthy users are less sensitive in their technology refresh timing decisions.

    This makes intutive sense: in a recession, replacing consumer electronics is a relative luxury, leading the most price-sensitive buyers — overwhelmingly Android users — to push off replacing otherwise functional devices.

    Voila! A slight dip in the overwhelming volume of Android purchases appears to bolster both ASP and iOS share since iOS devices are almost universally more expensive. ASP might be a fine diagnostic metric in limited cases, but it always pays to develop a sense for the texture of your data! ↩ī¸Ž

  6. While the model of median device representation presented here is hand-wavey, I've cross-referenced the results with Google's internal analysis and the model holds up OK. Native app and SDK developers with global reach can use RAM, CPU speeds, and screen DPI to bucket devices into high/medium/low tiers, and RAM has historically been a key signifier of device performance overall. Indeed, some large sites I've worked with have built device model databases to bucket traffic based on system RAM to decide which versions/features to serve.

    Our ASP + age model could be confounded by many factors: ASPs deviating more than expected year-on-year, non-linearities in price/performance (e.g., due to a global chip shortage), or sudden changes in device replacement rates. Assuming these factors hold roughly steady, so will the model's utility. Always gut-check the results of a model! Caveat emptor.

    In our relatively steady-state, 90+% of active smartphones were sold in the last 4 years. This makes some intuitive sense: batteries degrade, screens crack, and disks fill up. Replacement happens.

    If our napkin-back modelling has any bias, it's slightly too generous about device prices (too high) and age (too young) given the outsized historical influence of wealthier markets achieving smartphone saturation ahead of emerging one. Given that a new baseline is being set in an environment where we know the bottom-end will start to rise in terms of performance over the next few years (instead of staying stagnant), this bias, while hard to quantify, doesn't seem deeply problematic. ↩ī¸Ž

  7. While process nodes are starting to move forward in a hopeful way, the overall design tradeoffs of major Android-ecosystem SoC vendors are not. Recently updated designs for budget parts finally get out-of-order dispatch and better parallelism.

    Sadly, the feeble caches they're mated to ensure whatever gains are unlocked by better microarchitectures, plus higher frequencies, go largely to waste. It's not much good to clock a core faster or give it more ability to dispatch ops in if it's only spinning up to a high power state to stall on main memory, 100s of cycles away.

    Apple consistently wins benchmarks one both perf and power by throwing die space at the ARBs and caches to feed their hungry, hungry cores (among other improvements to core design). Any vendor with an ARM Architectural License can play this game, but either Qualcomm et al. are too stingy to fork out $200K (or $800K?) for a full license to customise the core of the thing they market, or they don't want to spend more to fund the extra design time or increase mm^2 of die space (both of which might cut into margins).

    Whatever the reason, these designs are (literally) hot garbage, dollar-for-dollar, compared to the price/performance of equivalent-generation Apple designs. It's a trickle-down digital divide, on a loop. The folks who suffer are end-users who can't access services and information they need quickly because developers in the privilege bubble remain insulated by wealth from all of these effects, oblivious to the ways that a market failure several layers down the stack keeps them ignorant of their malign impacts on users. ↩ī¸Ž ↩ī¸Ž

Election Season 2020, W3C TAG Edition

Update: Jeffery's own post discussing his run is now up and I recommend reading it if you're in a position to cast a vote in the election.

Two years ago, as I announced that I wouldn't be standing for a fourth term on the W3C's Technical Architecture Group, I wrote that:

Having achieved much of what I hoped when running for the TAG six years ago, it's time for fresh perspectives.

There was no better replacement I could hope for than Alice Boxhall. The W3C membership agreed and elected her to a two year term. Her independence, technical rigour, and dedication to putting users and their needs first have served the TAG and the broader web community exceedingly well since. Sadly, she is not seeking re-election. The TAG will be poorer without her dedication, no-nonsense approach, and depth of knowledge.

TAG service is difficult, time consuming, and requires incredible focus to engage constructively in the controversies the TAG is requested to provide guidance about. The technical breadth required creates a sort of whiplash effect as TAG members consider low-level semantics and high-level design issues across nearly the entire web platform. That anyone serves more than one term is a minor miracle!

While I'm sad Alice isn't running again, my colleague Jeffrey Yasskin has thrown his hat into the ring in one of the most contested TAG elections of recent memory. There aren't many folks who have the temperament, experience, and track record that would make them natural TAG candidates, but Jeffrey is in that rare group and I hope that I can convince you he deserves you organisation's first-choice vote.

Jeffrey and I have worked closely over the past 5 years on challenging projects that have pushed the boundaries of the web's potential forward, but which also posed large risks if done poorly. It was Jeff's involvement that lead to the "chooser model" for several device access APIs, starting with Web Bluetooth.

The explicit chooser Jeffrey advocated gives users control over which devices to connect to, nimbly addressing a difficult impasse.

This model has since been pressed into service multiple times, notably for Chrome's mediation of the Web USB, Web HID, and Web Serial APIs. The introduction of a novel style of chooser solved thorny problems.

Previously, security and feature teams were unable to find a compromise that provided sufficient transparency and control while mitigating abuses inherent in blanket prompts. Jeffrey's innovation created space for compromise and, when combined with other abuse mitigations he helped develop for Bluetooth, was enough to surmount the objections levelled against proposals for (over)broad grants.

Choosers have unlocked educational programming without heavyweight tools, sparked a renaissance in utilities that work everywhere, and is contributing to a reduction in downloads of potentially insecure native programs for controlling IoT devices.

Choosers at work in Chrome's WebUSB implementation
Choosers at work in Chrome's WebUSB implementation

There's so much more to talk about in Jeffrey's background and approach to facilitating structural solutions to thorny problems — from his work to (finally) develop a grounded privacy threat model as the TAG repeatedly requested, to his stewardship of the Web Packaging work the TAG started in 2013 to his deep experience in standards at nearly every level of the web stack — but this post is already too long.

There are a lot of great candidates running for this year's four open seats on the TAG, and while I can imagine many of them serving admirably, not least of all Tess O'Connor and Sanghwan Moon who have been doing great work, there's nobody I'd rather see earn your first-choice vote than Jeffrey Yasskin. He's not one to seek the spotlight for himself, but having worked closely with him, I can assure you that, if elected, he'll earn your trust the way he has earned mine.

Resize-Resilient `content-visibility` Fixes

Update: After hitting a bug related to initial rendering on Android, I've updated the code here and in the snippet to be resilient to browsers deciding (wrongly) to skip rendering the first <article> in the document.

Update, The Second: After holiday explorations, it turns out that one can, indeed, use contain-intrinsic-size on elements that aren't laid out yet. Contra previous advice, you absolutely should do that, even if you don't know natural widths yet. Assuming the browser will calculate a width for an <article> once finally laid out, there's no harm in reserving a narrow but tall place-holder. Code below has been updated to reflect this.

The last post on avoiding rendering work for content out of the viewport with content-visibility included a partial solution for how to prevent jumpy scrollbars. This approach had a few drawbacks:

Much of this was pointed out (if obliquely) in a PR to the ECMAScript spec. One alternative is a "place-keeper" element that grows with rendered content to keep elements that eventually disappear from perturbing scroll position. This was conceptually hacky and I couldn't get it working well.

What should happen if an element in the flow changes its width or height in reaction to the browser lazy-loading content? And what if the scroll direction is up, rather than down?

For a mercifully brief while I also played with absolutely positioning elements on reveal and moving them later. This also smelled, but it got me playing with ResizeObservers. After sleeping on the problem, a better answer presented itself: leave the flow alone and use IntersectionObservers and ResizeObservers to reserve vertical space using CSS's new contain-intrinsic-size property once elements have been laid out.

I'd dismissed contain-intrinsic-size for use in the base stylesheet because it's impossible to know widths and heights to reserve. IntersectionObservers and ResizeObservers, however, guarantee that we can know the bounding boxes of the elements cheaply (post layout) and use the sizing info they provide to reserve space should the system decide to stop laying out elements (setting their height to 0) when they leave the viewport. We can still use contain-intrinsic-size, however, to reserve a conservative "place-holder" for a few purposes:

  1. If our element would be content sized, we can give it a narrow intrinsic-size width; the browser will still give us the natural width upon first layout.
  2. Reserving some height for each element creates space for content-visibilitys internal Intersection Observer to interact with. If we don't provide a height (in this case, 500px as a guess), fast scrolls will encounter "bunched" elements upon hitting the bottom, and content-visiblity: auto will respond by laying out all of the elements at once, which isn't what we want. Creating space means we are more likely to lay them out more granularly.

The snippet below works around a Chrome bug with applying content-visibility: auto; to all <articles>, forcing initial paint of the first, then allowing it to later be elided. Observers update sizes in reaction to layouts or browser resizing, updating place-holder sizes while allowing the natural flow to be used. Best of all, it acts a progressive enhancement:

<!-- Basic document structure -->
/* Workaround for Chrome bug, part 1
* Chunk rendering for all but the first article.
* Eventually, this selector will just be:
* body > main > article

body > main > article + article {
content-visibility: auto;
/* Add vertical space for each "phantom" */
contain-intrinsic-size: 10px 500px;
<!-- header elements -->
<article><!-- ... --></article>
<article><!-- ... --></article>
<article><!-- ... --></article>
<!-- ... -->
<!-- Inline, at the bottom of the document -->
<script type="module">
let eqIsh = (a, b, fuzz=2) => {
return (Math.abs(a - b) <= fuzz);

let rectNotEQ = (a, b) => {
return (!eqIsh(a.width, b.width) ||
!eqIsh(a.height, b.height));

// Keep a map of elements and the dimensions of
// their place-holders, re-setting the element's
// intrinsic size when we get updated measurements
// from observers.
let spaced = new WeakMap();

// Only call this when known cheap, post layout
let reserveSpace = (el, rect=el.getClientBoundingRect()) => {
let old = spaced.get(el);
// Set intrinsic size to prevent jumping on un-painting:
if (!old || rectNotEQ(old, rect)) {
spaced.set(el, rect);
`${rect.width}px ${rect.height}px`

let iObs = new IntersectionObserver(
(entries, o) => {
entries.forEach((entry) => {
// We don't care if the element is intersecting or
// has been laid out as our page structure ensures
// they'll get the right width.
{ rootMargin: "500px 0px 500px 0px" }

let rObs = new ResizeObserver(
(entries, o) => {
entries.forEach((entry) => {
reserveSpace(, entry.contentRect);

let articles =
document.querySelectorAll("body > main > article");

articles.forEach((el) => {

// Workaround for Chrome bug, part 2.
// Re-enable browser management of rendering for the
// first article after the first paint. Double-rAF
// to ensure we get called after a layout.
requestAnimationFrame(() => {
requestAnimationFrame(() => {

This solves most of our previous problems:

Debugging this wouldn't have been possible without help from Vladimir Levin and Una Kravets whose article has been an indispensable reference, laying out the pieces for me to eventually (too-slowly) cobble into a full solution.

`content-visiblity` Without Jittery Scrollbars

Update: After further investigation, an even better solution has presented itself, which is documented in the next post.

The new content-visibility CSS property finally allows browsers to intelligently decide to defer layout and rendering work for content that isn't on-screen. For pages with large DOMs, this can be transformative.

In applications that might otherwise be tempted to adopt large piles of JS to manage viewports, it can be the difference between SPA monoliths and small progressive enhancements to HTML.

One challenge with naive application of content-visibility, though, is the way that it removes elements from the rendered tree once they leave the viewport -- particularly as you scroll downward. If the scroll position depends on elements above the currently viewable content "accordion scrollbars" can dance gleefully as content-visibility: auto does its thing.

To combat this effect, I've added some CSS to this site that optimistically renders the first article on a page but defers others. Given how long my articles are, it's a safe bet this won't usually trigger multiple renders for "above the fold" content on initial pageload (sorry not sorry).

This is coupled with an IntersectionObserver that to unfurls subsequesnt articles as you scroll down.

The snippets:

/* Defer rendering for the 2nd+ article */
body > main > *+* {
content-visibility: auto;
<script type="module">
let observer = new IntersectionObserver(
(entries, o) => {
entries.forEach((entry) => {
let el =;
// Not currently in intersection area.
if (entry.intersectionRatio == 0) {
// Trigger rendering for elements within
// scroll area that haven't already been
// marked.
if (!el.markedVisible) {
el.markedVisible = true;
// Set a rendering "skirt" 50px above
// and 100px below the main scroll area.
{ rootMargin: "50px 0px 100px 0px" }

let els =
document.querySelectorAll("body > main > *+*");
els.forEach((el) => { observer.observe(el); });

With this fix in place content will continue to appear at the bottom (ala "infinite scrolling") as you scroll down, but never vanish from the top or preturb your scroll position in unnatural ways.

Older Posts

Newer Posts