内容简介:Web performance isn’t going to save you in this crisis.But if you’re building a software product, chances are you’re relatively unaffected. And in this case, having a quick app is more important than ever. InternetAnd
Web performance isn’t going to save you in this crisis.
But if you’re building a software product, chances are you’re relatively unaffected. And in this case, having a quick app is more important than ever. Internet is slowing down due to increased demand, and people are holding on to their phones for longer – so if your app is slow, your users will be affected by it.
And slow app means worse business .
Today, let’s take a look at Notion – an advanced note-taking web app. Notion is a great product, but one thing customers complain about is its startup time:
Notion app and my laptop OS take the same time to boot.
Beautiful application but it's tooooo slow:pensive:
— Aniket Mandle (@aniketmandle11) May 2, 2020Notion shipped tremendous speed enhancements recently , but there’s still a lot of space to improve. Let’s reverse-engineer it and see what else can be optimized!
How Notion loads
Notion is a React web app embedded into a native shell. This means its “startup time” is largely the “loading time of the web part”.
To be concrete.On the desktop, the Notion app is a web app wrapped with Electron. On mobile, from my understanding, the Notion app runs both React Native parts (likely responsible for some mobile experiences) and web parts (likely responsible for the overall editing UI).
Due to (apparently) HTTPS certificate pinning in the Android Notion app, I wasn’t able to verify whether the mobile app runs the same bundles the desktop app does. But even if the bundles are different, issues they are having are, likely, similar.
To understand how the web part is loading, let’s create a public Notion page :
and run a WebPageTest audit over it. (This works because public pages run the same code the whole app runs.)
Sidenote: WebPageTest is an advanced performance testing tool.
WebPageTest returns a lot of useful information, but the most interesting part is the loading waterfall :
Woah, that’s a lot of info. What’s going on there?
Here’s what’s going on:
- You open the page. The page loads a few stylesheets – and two JavaScript bundles,
vendor
andapp
. - Once both bundles are loaded, they start to execute – and spend a whole second doing so.
- Once the app has been initialized, it starts sending API requests for page data, loading analytics...
- and executing more code...
-
until, at 5.6 seconds, the first paint happens:
...but it’s just a spinner.
-
At 6.2 seconds, the page content gets actually rendered .
It takes a couple more seconds to finish loading all the hero image.
6.2 seconds for a desktop computer is quite a lot. However, with a medium-tier phone like Nexus 5, this time increases to 12.6 seconds . Here’s how that feels:
Let’s see how we can improve it.
Cost of JavaScript
When one talks about “loading speed”, one typically means networking performance . From the networking standpoint, Notion is doing well: they’re using HTTP/2, they’re gzipping files, and they’re using Cloudflare as a proxying CDN.
However, another part of “loading speed” that people talk less about is processing performance . All downloaded resources have a processing cost: gzip archives need to be decompressed; images need to be decoded; JS needs to be executed.
Unlike networking performance, processing performance doesn’t improve with better networks – it’s only as fast as the user’s CPU is. And user’s CPUs in phones – especially Android phones – are bad :
For Notion, processing performance is even more significant. It’s easy to avoid networking costs by caching network resources in the app. But processing costs are paid every single time the app starts – which means a phone user may see a 10-second splash screen multiple times a day.
On our test Nexus 5, executing vendor
and app
bundles takes 4.9 seconds. This whole time, the page – and the app – stay non-interactive and empty:
What’s happening there? WebPageTest doesn’t record JS traces, but if we go to DevTools and run a local audit, we’ll see this:
First, the vendor
bundle is being compiled (for 0.4s). Second, the app
bundle is being compiled (for 1.2s). Third, both bundles start executing – and spend 3.3s doing so.
So how can we reduce that time?
Defer JS execution
Let’s take a look at the bundle execution phase. What are all these functions?
Turns out that’s bundle initialization:
-
Functions with four-character names , like
bkwR
orCycz
, are application modules.Sidenote: Four-character IDs (like
bkwR
orCycz
) are used whenoptimization.moduleIds: 'hashed'
orHashedModuleIdsPlugin
is enabled.| | When webpack builds a bundle, it wraps each module with a function – and assigns it an ID. That ID becomes the function name. In the bundle, this looks as follows: | | Before: | |
js | import formatDate from './formatDate.js'; | | // ... |
| | After: | |js | fOpr: function(module, __webpack_exports__, __webpack_require__) { | "use strict"; | | __webpack_require__.r(__webpack_exports__); | | var _formatDate__WEBPACK_IMPORTED_MODULE_0__ = | __webpack_require__("xN6P"); | | // ... | }, |
-
And the
s
function is actually__webpack_require__
.__webpack_require__
is the webpack’s internal function that it uses for requiring modules. Whenever you write animport
, webpack transforms it to__webpack_require__()
.
Bundle initialization takes so much time because it executes all the modules. Each module may take just a few milliseconds to execute, but with Notion’s 1100+ modules, this adds up.
The only way to fix it is to execute fewer modules upfront.
Use code splitting
The best way to improve the startup time is to code-split away some features that’re not needed immediately. In webpack, this is done using import()
:
// Before <Button onClick={openModal} /> // After <Button onClick={() => import('./Modal').then(m => m.openModal())} />
Code splitting is the best first optimization you can do.It brings huge performance benefits: after doing code splitting, Tinder reported a 60% decrease in the complete load time ; and our client, Framer, managed to cut 40-45% off the CPU Idle time .
There are several common approaches to code splitting :
- splitting the bundle by pages,
- splitting away below-the-fold code,
- and splitting away conditional content (any dynamic UIs that are not visible immediately)
The Notion app doesn’t have pages, and code-splitting below-the-fold content is hard because pages are very dynamic. This means the only useful approach is code-splitting conditional content. The following parts may be good candidates for that:
- Settings , Import , Trash – all UIs that are rarely used
- Sidebar , Share , Page options – all UIs that are frequently used but are not needed immediately when the app starts. Those could be preloaded and initialized right after the app starts
- Heavy page blocks. Some page blocks are quite heavy – e.g. the Code block supports highlighting 68 languages, which bundles 120+ minified KBs of language definitions from Prism.js. Notion seems to be code-splitting some blocks already (e.g., Math equation ), but it might make sense to extend it to others as well.
Check that module concatenation is working
In webpack, the module concatenation feature is responsible for merging multiple small ES modules into one large. This reduces the module processing overhead and makes removing unused code more effective.
To confirm that module concatenation is working:
- check that you’re not transpiling ES modules to CommonJS with Babel.
@babel/preset-env
doesn’t transpile ES modules into CommonJS - check that
optimization.concatenateModules
is not disabled explicitly - run the production webpack build with
--display-optimization-bailout
to see if there are any cases where module concatenation bails out
Fun fact.Remember that all imports are transformed into the __webpack_require__
function?
Well, what happens when the same function is called 1100 times throughout initialization? Right, it becomes a hot path taking 26.8% of the total time:
( s
is the minified name of __webpack_require__
.)
Unfortunately, apart from concatenating more modules, there’s not much to optimize there .
Try the lazy
option of Babel’s plugin-transform-modules-commonjs
Note:this suggestion relies on disabling module concatenation. Because of that, it’s incompatible with the previous one.
@babel/plugin-transform-modules-commonjs
is an official Babel plugin that transforms ES imports into CommonJS require()
s:
// Before import formatDate from './formatDate.js'; export function getToday() { return formatDate(new Date()); } // After const formatDate = require('./formatDate.js'); exports.getToday = function getToday() { return formatDate(new Date()); };
And with its lazy
option enabled, it also inlines all require
s right into where they’re used:
// After, with `lazy: (path) => true`, simplified exports.getToday = function getToday() { return require('./formatDate.js')(new Date()); };
Thanks to this transformation, if the getToday()
function is never called, ./formatDate.js
is never imported! And we don’t pay the import cost.
There are a few drawbacks, however:
- Switching the existing codebase to
lazy
might be tricky. Some modules may rely on side effects from other modules, which we’re delaying here. Also, the plugin docs warn that thelazy
option breaks cyclic dependencies -
Switching to CommonJS modules disables tree shaking in webpack . This means some unused code may be kept in the bundle
Sidenote: This doesn’t affect webpack 5, as it will add support for CommonJS tree shaking .
- Switching to CommonJS modules disables module concatenation . This means the module processing overhead will be higher
These drawbacks make this option riskier compared to others – but if it plays right, its benefits might far outweigh its costs.
How many modules could be deferred like this?Chrome DevTools let us find an easy answer. Open any JS-heavy page (e.g. the Notion one ), go to DevTools, press Ctrl+Shift+P (Windows) or ⌘⇧P (macOS), type in “start coverage”, and press Enter. The page will reload, and you’ll see how much code was executed in the initial render.
In Notion, 39% of the vendor bundle and 61% of the app bundle are unused after the page renders:
Remove unused JS code
Let’s take a look at the bundle initialization trace again:
A significant part here is “Compile Script” (parts 1 and 2), which takes 1.6s in total. What is that?
V8 (Chrome’s JS engine), just like other JS engines, uses just-in-time compilation to run JavaScript. This means all code it executes has to be compiled into machine code first.
And the more code there is, the more time it takes to compile it. In 2018, on average, V8 was spending 10-30% of total execution time in parsing and compiling JavaScript. In our case, the compilation step takes 1.6s out of a total of 4.9s – a whopping 32%.
The only way to reduce the compilation time is to serve less JavaScript.
Sidenote: Another great approach would be to precompile JavaScript into machine code – and avoid parsing costs altogether by running compiled JavaScript. However, this is not currently possible.
Use code splitting
Yes, this again. By code-splitting unused functionality, you not onlyreduce bundle init time, but also decrease the compilation time. The less JS code there is, the faster it compiles.
Check outthe previous section on code splittingwhere we’ve talked about common code-splitting approaches & how Notion could benefit from it.
Remove unused vendor code
, when a page loads, almost 40% of the Notion’s vendor
bundle stay unused:
Some of that code will be needed later when the user does something in the app. But how much?
Notion doesn’t publish source maps, which means we can’t use source-map-explorer
to explore the bundle and see the largest modules. However, we could still guess libraries from their minified source – by looking at non-minified strings & searching for them in GitHub.
Based on my analysis, here’re the 10 largest modules in the vendor
bundle:
fingerprintjs2 moment-timezone chroma-js tinymce diff-match-patch amplitude-js lodash libphonenumber-js/metadata.min.json react-dom moment
Sidenote:This list does not include libraries that are composed of multiple small files.
For example, the bundle also includes core-js
, which occupies 154 KB but consists of 300+ small files.
Out of all these modules, the most significant and easy to optimize ones are moment
, lodash
and libphonenumber-js
.
moment
, a JS library for manipulating dates, is bundling 160+ minified KBs of localization files. Given than Notion is only available in English, this is hardly needed.
What can one do here?
- First, drop unused
moment
locales usingmoment-locales-webpack-plugin
. -
Second, consider switching from
moment
todate-fns
. Unlike withmoment
, when you’re usingdate-fns
, you’re importing only specific date manipulation methods you need. So if you only useaddDays(date, 5)
, you won’t end up bundling the date parser.Sidenote: You-Dont-Meed-Momentjs : List of functions which you can use to replace moment.js
lodash
, a set of data manipulation utilities, bundles 300+ functions for working with data. That’s too much – from what I’ve seen, apps typically use 5-30 of those methods at most.
The easiest way to drop unused methods is to use babel-plugin-lodash
. Apart from that, lodash-webpack-plugin
supports removing some lodash features (like caching or Unicode support) from inside these methods.
libphonenumber-js
, a library for parsing and formatting phone numbers, bundles an 81 KB JSON file with phone number metadata.
I can’t see any places where phone numbers are used, so it’s likely this library supports a single use case somewhere deep in the Notion UI. It’d be great to replace it with another library or custom code – and drop the whole dependency.
Remove polyfills
Another major dependency present in the vendor
bundle is polyfills from the core-js
library :
There are two problems with it.
It’s unnecessary.We’re testing Notion in Chrome 81, which supports all the modern JS features. However, the bundle still includes polyfills for Symbol
, Object.assign
, and many other methods. These polyfills have to be downloaded, parsed, and compiled – all for nothing.
This also affects Notion apps. In the desktop app (and probably in the mobile one as well), the JS engine version is modern, fixed, and well-known. There’s zero chance Symbol
or Object.assign
would be absent there – however, the app still downloads the same polyfills.
What should we do instead? Ship polyfills for older browsers, but skip them for modern ones. See “How to load polyfills only when needed” for a few ways to do this.
It’s bundled multiple times.The vendor
bundle includes the core-js
copyright 3 times. Each time, the copyright is identical, but is shipped in a different module and with different dependencies:
This means core-js
itself is bundled 3 times. But why? Let’s dig deeper.
In a non-minified form, the module with the copyright looks like this :
var core = require('./_core'); var global = require('./_global'); var SHARED = '__core-js_shared__'; var store = global[SHARED] || (global[SHARED] = {}); (module.exports = function (key, value) { return store[key] || (store[key] = value !== undefined ? value : {}); })('versions', []).push({ version: core.version, mode: require('./_library') ? 'pure' : 'global', copyright: '© 2019 Denis Pushkarev (zloirock.ru)', });
Here, we have two bits that describe the library:
-
var core = require('./_core'); core.version
for the library version, and -
require('./_library') ? 'pure' : 'global'
for the library mode
In the minified code, that corresponds to:
var r=n(<MODULE_ID>);r.version n(<MODULE_ID>)?"pure":"global"
If we follow these module IDs in the bundle, we’ll see this:
Woah. This means these three versions of core-js
are:
-
2.6.9
in theglobal
mode, -
2.6.11
in theglobal
mode, and -
2.6.11
in thepure
mode
Turns out, this is a common issue . This happens when your app depends on one version of core-js
, but some of your dependencies depend on another.
How to solve it? Run yarn why core-js
to figure out what depends on the remaining two versions. And either remove/reconfigure dependencies that bundle extra core-js
versions; or deduplicate all three versions into one using webpack’s resolve.alias
:
Okay, let’s talk about bundle duplicates.
A common issue in JS bundles is duplicated dependencies. E.g., lodash, core-js, some polyfill libs are frequently bundled multiple times.
Here’s how to detect and solve this issue: a thread ⬇ pic.twitter.com/sF3h486v3U
— Ivan Akulov (@iamakulov) May 18, 2020Optimize the loading waterfall
Let’s take another look at how Notion is loading :
A few things are striking the attention here:
loadPageChunk
Here’s how to optimize that.
Defer third parties
Third parties are scripts from other domains we add for ads, analytics, or similar functionality. Business-wise, they’re both useful and problematic:
- Useful: we add third parties for a concrete business value (e.g., analyzing how users use the app) that we won’t get otherwise
- Problematic: third parties typically hurt loading performance which negatively affects conversion and user satisfaction
Sidenote: Harry Roberts has a great talk about third parties: It’s my (third) party, and I’ll cry if I want to
In Notion, third parties (Intercom, Segment, and Amplitude) hurt performance by executing JavaScript and blocking the main thread when it’s most needed – when the app is still initializing. If we block them in WebPageTest, our Nexus 5 will render the content a whole second earlier :
In the real life, we can’t simply remove all the Notion third parties. But we can defer them – like this:
// Before async function installThirdParties() { if (state.isIntercomEnabled) intercom.installIntercom(); if (state.isSegmentEnabled) segment.installSegment(); if (state.isAmplitudeEnabled) amplitude.installAmplitude(); } // After async function installThirdParties() { setTimeout(() => { if (state.isIntercomEnabled) intercom.installIntercom(); if (state.isSegmentEnabled) segment.installSegment(); if (state.isAmplitudeEnabled) amplitude.installAmplitude(); }, 15 * 1000); }
This would make sure they are not loaded until the app has fully initialized.
setTimeout
vs requestIdleCallback
vs events. setTimeout
is not the best approach (hard-coding the timeout is hacky), but it’s good enough.
The best approach would be to listen for some kind of a “page fully rendered” in-app event, but I’m not sure whether Notion has one.
requestIdleCallback
might sound like the perfect tool for the job, but it’s not. In my tests in Chromium, it triggers too early – merely 60 ms after the main thread becomes idle.
Loading analytics on interaction.Another great approach to defer analytics is to avoid loading it until the first user’s interaction – the first click or tap.
However, note that this makes analytics invisible for synthetic tests (like Lighthouse or PageSpeed Insights). To measure the real JavaScript cost for users, you should install a Real User Monitoring library – e.g. LUX from SpeedCurve or Browser Insights from Cloudflare .
Preload API data
In Notion, before the page is rendered, the browser has to send 9 requests to the API:
Each request may take from 70 ms (in case of a cable connection ) to 300-500 ms (in case of a 4G connection and a medium-tier phone ). And some of these requests seem sequential – they aren’t sent before previous requests complete.
This means slow API requests can easily result in significant latency. In my tests, removing this latency makes the page render 10% faster .
But how can we remove the latency in the real app?
Inline page data into the HTML.The best approach would be to calculate the API data on the server side – and include it directly into the HTML response. E.g., like this:
app.get('*', (req, res) => { /* ... */ // Send the bundles so the browser can start loading them res.write(` <div id="notion-app"></div> <script src="/vendors-2b1c131a5683b1af62d9.js" defer></script> <script src="/app-c87b8b1572429828e701.js" defer></script> `); // Send the initial state when it’s ready const stateJson = await getStateAsJsonObject(); res.write(` <script> window.__INITIAL_STATE__ = JSON.parse(${stateString}) </script> `); })
Sidenote:Make sure to:
a) encode data as JSON for best performance ;
b) escape data with jsesc
( json: true, isScriptContext: true
) to avoid XSS attacks.
Also, note that bundles have the defer
attribute. We need it to execute bundles after the __INITIAL_STATE__
script.
With this approach, the app won’t need to wait for API responses. It will retrieve the initial state from the window
and start rendering immediately.
Cloudflare workers.Notion uses Cloudflare as a CDN provider. If Notion’s HTML pages are static (e.g., they’re served by AWS S3), Cloudflare workers might be useful instead.
With Cloudflare workers, you can intercept the page, fetch dynamic data straight from the CDN worker, and append the data into the end of the page. See:
- Streams documentation for details on how to transform the response on the go
- Streaming recipes and Fast Google Fonts for some examples of how this can be done
Inline a script to prefetch page data.Another approach is to write an inline script that will request the data ahead of time:
<div id="notion-app"></div> <script> fetchAnalytics(); fetchExperiments(); fetchPageChunk(); function fetchAnalytics() { window._analyticsSettings = fetch( '/api/v3/getUserAnalyticsSettings', { method: 'POST', body: '{"platform": "web"}', } ).then((response) => response.json()); } async function fetchExperiments() { /* ... */ } async function fetchPageChunk() { /* ... */ } </script> <script src="/vendors-2b1c131a5683b1af62d9.js"></script> <script src="/app-c87b8b1572429828e701.js"></script>
The app can then simply await
on window._analyticsSettings
(and similar promises). If the data is loaded by that time, the app will get it near-immediately.
The important bit: the script should start sending requests as soon as possible. This will increase the chance that responses will arrive – and be handled – while the bundles are still loading and the main thread is idle.
Other stuff
Optimizations above should bring the most benefits. But there are a few other things that are worth paying attention to.
Cache-Control
on responses
Notion doesn’t set the Cache-Control
header on its responses. This doesn’t disable caching – but means each browser would cache the response differently . This could lead to some unexpected client-side bugs.
To avoid this, set the proper Cache-Control
header on bundle assets and API responses:
This is how you typically do caching BTW:
For API responses (like /api/user): prevent caching
→ Cache-Control: max-age=0, no-store
For hashed assets (like /static/bundle-ab3f67.js): cache for as long as possible
→ Cache-Control: max-age=31556952, immutable https://t.co/BHetokxuJO
— Ivan Akulov (@iamakulov) May 11, 2020Loading skeleton
The Notion app has a spinner that’s shown while the page is loading:
The spinner helps to signify that “something is loading”. However, sometimes, the spinner actually worsens the perceived performance . Users see the spinner and pay attention to the fact that something is loading – which makes the app feel slower.
What could be done instead is showing a skeleton of the UI :
It’s small enough to be inlineable, and it prepares the user for the actual UI.
Summing up
So, how much time can all these optimizations save us?
-
Let’s assume that we manage to shave 30% off the vendor bundle byremoving unused dependenciesandnot shipping polyfills. Additionally, let’s assume we manage to20% of the main bundle.
It’s hard to say how much that will reduce the compilation and execution time – the effect can be from 10% to 50% – but let’s stick with 25%. On our Nexus 5, total JS timewithout third partieswas 6.4s – so 25% of it would be 1.6s.
- Preloading the API datacan reduce time to contentful paint by 10%. With Nexus 5, that’ll be 10% of 12.6s = 1.3s.
- Anddeferring third partiescan save us another second.
In total, based on this (very rough) calculation,we save 3.9 out of 12.6 seconds – a 30% improvementjust by tuning some configs and deferring some loading. And this is after great speed improvements the Notion team has already done.
It turns out, almost every app has low-hanging fruits that could be implemented just by tuning the bundler configuration and doing a few precise code changes. So here’s the easiest way to find and pick them . And if you read this far and liked this case study, consider spreading the word about it:
Finally published the article I’ve been working on for the past 3 weeks!
:memo: Case study: Analyzing Notion app performance
Or how to make a React app load ~30% faster – just by tuning some configs and delaying some scripts.
→ https://t.co/3rBsup6dCg pic.twitter.com/DN5fVvxBRZ
— Ivan Akulov (@iamakulov) May 19, 2020Thanks to Radion Chernyakov, Semyon Muravyov, Victor Kolb, Nikolay Kost for their draft reviews and helpful suggestions.
以上就是本文的全部内容,希望对大家的学习有所帮助,也希望大家多多支持 码农网
猜你喜欢:本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们。
Visual LISP程序设计
李学志 / 清华大学 / 2006-5 / 29.00元
本书系统地介绍了AutoCAD最新版本(2006)的Visual LISP程序设计技术。全书共分13章。前3章介绍AutoLISP语言的基础知识,第4章介绍Visual LISP的开发环境,第5~7章介绍程序的编辑、调试和设计的方法与技巧,第8章介绍如何定义新的AutoCAD命令及创建图层、线型、文字样式、剖面线、尺寸标注等各种AutoCAD对象,以及如何实现参数化图形设计的方法和技术,第9章介绍......一起来看看 《Visual LISP程序设计》 这本书的介绍吧!