探究 LightHouse 工作流程

大厂技术 坚持周更 精选好文

本文为来自 字节教育-智能学习-前端团队的文章,已授权 ELab 发布。

智能学习前端团队自创立以来,团队专注于打破大众对教育的刻板印象,突破固有的教学思维,攻破各类教学屏障。旨在为每一位学生制定最合适的学习方案,予以因材施教,使优质教育随‘触’可达。

什么是Lighthouse

Lighthouse analyzes web apps and web pages, collecting modern performance metrics and insights on developer best practices.

使用方式

原理结构[1]

Gathering

Driver 驱动

通过 Chrome Debugging Protocol 和 Puppeteer[2](提供无头浏览器环境模拟页面操作) /进行交互。

Chrome Debugging Protocol(CDP)

Chrome DevTools 协议允许使用工具来检测、检查、调试和分析 Chromium、Chrome 和其他基于 Blink 的浏览器。 在Chrome扩展中,Chrome protocol 利用 chrome.debuggerApi 通过 WebSocket[3]来建立连接。

Instrumentation 分为多个 Domains(DOM, Debugger, Network 等)。每个 Domain 定义了许多它支持的命令和它生成的事件。命令和事件都是固定结构的序列化 JSON 对象。

CDPDomains,红色为实验性

Domain 必须 enable后才可以发出事件。一旦启用enable,它们将刷新表示状态的所有事件。因此,网络事件仅在enable后才会发出。所有协议代理解析 enable 的回调。比如:

// will NOT work
driver.defaultSession.sendCommand('Security.enable').then(_ => {
driver.defaultSession.on('Security.securityStateChanged', state => { /* ... */ });
})

// WILL work! happy happy. :)
driver.defaultSession.on('Security.securityStateChanged', state => { /* ... */ }); // event binding is synchronous
driver.defaultSession.sendCommand('Security.enable');

配置

passes

passes 属性控制如何加载请求的 URL,以及在加载时收集哪些关于页面的信息。pass 数组中的每个条目代表页面的一次加载,

每个 pass都定义了基本设置,例如等待页面加载多长时间、是否记录 trace 文件。此外,每次传递都定义了要使用的gatherer列表。gatherer可以从页面中读取信息以生成artifacts,稍后 Audits 使用这些artifacts提供 Lighthouse 报告。

具体的 pass 配置示例:

{
passes: [{
passName: 'fastPass',
atherers: ['fast-gatherer'],
},
{
passName: 'slowPass',
recordTrace: true,
useThrottling: true,
networkQuietThresholdMs: 5000,
gatherers: ['slow-gatherer'],
}]
}

Gatherers 采集器

决定在页面加载过程中采集哪些信息,将采集的信息输出为 artifacts。使用 Driver 采集页面信息。用--gather-mode指令运行可以获得3个采集产物:

  1. artifacts.json: 所有采集器的输出。

  1. defaultPass.trace.json: 大多数性能指标。可以在DevTools性能面板中查看。

  1. defaultPass.devtoolslog.json: DevTools Protocol[5]事件的日志。

每一个 gatherer继承自相同的基类 Gatherer,基类 Gatherer定义了传递生命周期的n个方法。gathererartifacts是生命周期方法返回的最后一个未定义值,所有方法都可以直接返回artifacts或返回解析为该值的 Promise。子类只需实现生命周期方法即可。

比如用于js覆盖率的 gatherer:

该实例实现了 startInstrumentation 、stopInstrumentation、getArtifact 3个生命周期方法,其

class JsUsage extends FRGatherer {
meta = {
supportedModes: ['snapshot', 'timespan', 'navigation'],
};

constructor {
super;
this._scriptUsages = ;
}
async startInstrumentation(context) {
const session = context.driver.defaultSession;
await session.sendCommand('Profiler.enable');
await session.sendCommand('Profiler.startPreciseCoverage', {detailed: false});
}


async stopInstrumentation(context) {
const session = context.driver.defaultSession;
const coverageResponse = await session.sendCommand('Profiler.takePreciseCoverage');
this._scriptUsages = coverageResponse.result;
await session.sendCommand('Profiler.stopPreciseCoverage');
await session.sendCommand('Profiler.disable');
}

asyncgetArtifact {
const usageByScriptId = {};
for (const scriptUsage of this._scriptUsages) {
if (scriptUsage.url === '' || scriptUsage.url === '_lighthouse-eval.js') {
continue;
}
usageByScriptId[scriptUsage.scriptId] = scriptUsage;
}
return usageByScriptId;
}
}

class FRGatherer { meta = { supportedModes : }; //在任意时间段内开始观察页面 startInstrumentation ( passContext ) { } //Sensitive开始观察页面 startSensitiveInstrumentation ( passContext ) { } //Sensitive停止观察页面的方法 stopSensitiveInstrumentation ( passContext ) { } //在任意时间段内结束观察页面 stopInstrumentation ( passContext ) { } //收集有关页面的结果 getArtifact ( passContext ) { } /** * Legacy */ getname {} async beforePass ( passContext ) {} pass ( passContext ) { } async afterPass ( passContext, loadData ) {}

当 pass 中定义的所有 gatherers 运行完后,就会生成一个中间产物 artifacts,此后 Lighthouse 就可以断开与浏览器的连接,只使用 artifacts 进行后续的分析。

Trace 链路追踪

core/lib/tracehouse/trace-processor.js提供了链路到更有意义对象的转换。每个原始trace event[6]都具有以微秒为单位增长的时间戳、线程ID、进程ID、持续时间以及其他适用的元数据属性(比如事件类型、任务名称、帧等)

Example Trace Event

{
'pid': 41904, // process ID
'tid': 1295, // thread ID
'ts': 1676836141, // timestamp in microseconds
'ph': 'X', // trace event type
'cat': 'toplevel', // trace category from which this event came
'name': 'MessageLoop::RunTask', // relatively human-readable description of the trace event
'dur': 64, // duration of the task in microseconds
'args': {}, // contains additional data such as frame when applicable
}

Processed trace

Processed trace 可识别关键时刻的 trace 事件((navigation start, FCP, LCP, DCL, trace end 等),并过滤出主进程和主线程事件的视图。

{
processEvents: [/* all trace events in the main process */],
mainThreadEvents: [/* all trace events on the main thread */],
timings: {
timeOrigin: 0, // timeOrigin is always 0
msfirstContentfulPaint: 150, // firstContentfulPaint time in ms after time origin
/* other key moments */
traceEnd: 16420, // traceEnd time in ms after time origin
},
timestamps: {
timeOrigin: 623000000, // timeOrigin timestamp in microseconds, marks the start of the navigation of interest
firstContentfulPaint: 623150000, // firstContentfulPaint timestamp in microseconds
/* other key moments */
traceEnd: 639420000, // traceEnd timestamp in microseconds
},
}

实现

  1. Connecting to browser

  1. Resetting state with about:blank

  1. Navigate to about:blank

  1. Benchmarking machine

  1. Initializing…

  1. Preparing target for navigation mode

  1. Running defaultPass pass

  1. Resetting state with about:blank

  1. Navigate to about:blank

  1. Preparing target for navigation

  1. Cleaning origin data

  1. Cleaning browser cache

  1. Preparing network conditions

  1. Beginning devtoolsLog and trace

  1. Loading page & waiting for onload

  1. Navigating to https:XXX

  1. Gathering in-page: XXXXXXXX. (xN)

  1. Gathering trace

  1. Gathering devtoolsLog & network records

  1. Gathering XXX (xN)

begin;
|
→ runLighthouse;
|
→ legacyNavigation;


async function legacyNavigation(url, flags = {}, configJSON, userConnection) {
//...
const connection = userConnection || new CriConnection(flags.port, flags.hostname);
const artifacts = await Runner.gather( => {
const requestedUrl = UrlUtils.normalizeUrl(url);
return Runner._gatherArtifactsFromBrowser(requestedUrl, options, connection);
}, options);
return Runner.audit(artifacts, options);
}

static async _gatherArtifactsFromBrowser(requestedUrl, runnerOpts, connection) {
//创建connection的Driver
const driver = runnerOpts.driverMock || new Driver(connection);
const gatherOpts = {
driver,
requestedUrl,
settings: runnerOpts.config.settings,
computedCache: runnerOpts.computedCache,
};
const artifacts = await GatherRunner.run(runnerOpts.config.passes, gatherOpts);
return artifacts;
}
/****** GatherRunner ****/
static async run(passConfigs, options) {


//1.Connecting to browser
//通过 Websocket 建立连接, 基于 Chrome Debugging Protocol 通信
// CDPSession 实例用于与 Chrome Devtools 协议的原生通信
await driver.connect;
// 在 devtools/extension 案例中,我们在尝试清除状态时仍不能在站点上
// 所以我们首先导航到 about:blank,然后应用我们的仿真和设置
// 2.Resetting state with about:blank & 3.Navigating to blankPage
await GatherRunner.loadBlank(driver);
// 4. Benchmarking machine
const baseArtifacts = await GatherRunner.initializeBaseArtifacts(options);

// ...processing benchmarkIndex

// 5. Initializing…
await GatherRunner.setupDriver(driver, options);

let isFirstPass = true;
// each pass
for (const passConfig of passConfigs) {
const passContext = {
gatherMode: 'navigation',
driver,
url: options.requestedUrl,
settings: options.settings,
passConfig,
baseArtifacts,
computedCache: options.computedCache,
LighthouseRunWarnings: baseArtifacts.LighthouseRunWarnings,
};
//Starting from about:blank, load the page and run gatherers for this pass.
const passResults = await GatherRunner.runPass(passContext);
Object.assign(artifacts, passResults.artifacts);

// If we encountered a pageLoadError, don't try to keep loading the page in future passes.
if (passResults.pageLoadError && passConfig.loadFailureMode === 'fatal') {
baseArtifacts.PageLoadError = passResults.pageLoadError;
break;
}

if (isFirstPass) {
await GatherRunner.populateBaseArtifacts(passContext);
isFirstPass = false;
}
}

await GatherRunner.disposeDriver(driver, options);
return finalizeArtifacts(baseArtifacts, artifacts);
} catch (err) {
// Clean up on error. Don't await so that the root error, not a disposal error, is shown.
GatherRunner.disposeDriver(driver, options);

throw err;
}
}

_connectToSocket(response) {
const url = response.webSocketDebuggerUrl;
this._pageId = response.id;

return new Promise((resolve, reject) => {
const ws = new WebSocket(url, {
perMessageDeflate: false,
});
ws.on('open', => {
this._ws = ws;
resolve;
});
ws.on('message', data => this.handleRawMessage(/** @type {string} */ (data)));
ws.on('close', this.dispose.bind(this));
ws.on('error', reject);
});
}



static async setupDriver(driver, options) {
//...
await GatherRunner.assertNoSameOriginServiceWorkerClients(session, options.requestedUrl);
// 6. Preparing target for navigation mode,通过为全局 API 或错误处理启用协议域、仿真和新文档处理程序,准备在导航模式下分析的目标。
await prepare.prepareTargetForNavigationMode(driver, options.settings);
}

static async runPass(passContext) {
//7. Running defaultPass pass

const gathererResults = {};
const {driver, passConfig} = passContext;

// 8.Resetting state with about:blank 9.Navigating to about:blankGo to about:blank
// set up

await GatherRunner.loadBlank(driver, passConfig.blankPage);
// 10.Preparing target for navigation ~ 13.Preparing network conditions
const {warnings} = await prepare.prepareTargetForInpidualNavigation(
driver.defaultSession,
passContext.settings,
{
requestor: passContext.url,
disableStorageReset: !passConfig.useThrottling,
disableThrottling: !passConfig.useThrottling,
blockedUrlPatterns: passConfig.blockedUrlPatterns,
}
);
// run `startInstrumentation /beforePass` on gatherers.
passContext.LighthouseRunWarnings.push(...warnings);
await GatherRunner.beforePass(passContext, gathererResults);

// 14.Beginning devtoolsLog and trace,
// await driver.beginDevtoolsLog; await driver.beginTrace(settings);
await GatherRunner.beginRecording(passContext);
//15.Loading page & waiting for onload ,16.Navigating to https:XXX
const {navigationError: possibleNavError} = await GatherRunner.loadPage(driver, passContext);
//17.Gathering in-page: XXXXXXXX,run `pass` on gatherers.
await GatherRunner.pass(passContext, gathererResults);
const loadData = await GatherRunner.endRecording(passContext);

//18.Gathering trace 19.Gathering devtoolsLog & network records
await emulation.clearThrottling(driver.defaultSession);

//process page error

// If no error, save devtoolsLog and trace.
GatherRunner._addLoadDataToBaseArtifacts(passContext, loadData, passConfig.passName);

// 20.Gathering XXX. Run `afterPass(stopInstrumentation -> getArtifact )` on gatherers and return collected artifacts.
await GatherRunner.afterPass(passContext, loadData, gathererResults);
const artifacts = GatherRunner.collectArtifacts(gathererResults);


return artifacts;
}

Auditing

Audits 审查器

配置

实现

  1. Analyzing and running audits

  1. Auditing: XXX

  1. Generating results...

 async function legacyNavigation(url, flags = {}, configJSON, userConnection) {
//...
return Runner.audit(artifacts, options);
}
static async audit(artifacts, options) {

//...
//1. Analyzing and running audits &2.Auditing: XXX
const auditResultsById = await Runner._runAudits(settings, config.audits, artifacts,
lighthouseRunWarnings, computedCache);

//3.Generating results...
if (artifacts.LighthouseRunWarnings) {
lighthouseRunWarnings.push(...artifacts.LighthouseRunWarnings);
}

//....

}

Report 报告

客户端根据生成 Audit 结果的 LHR.json (Lighthouse Result) 生成结果报告页。评分报告,它包含了性能(Performance),访问无障碍(Accessibility),最佳实践(Best Practice),搜索引擎优化(SEO),PWA(Progressive Web App)5 个部分,每一项下面又有若干小项(audit),还有详细诊断结果和优化建议,帮助开发者有针对性地进行优化。

例如:在 Lighthouse 8 中,性能得分由以下几项的得分按不同的权重相加而得:

Lighthouse8 中性能指标权重

如何确定指标分数

以性能评分[7]为例,一旦 Lighthouse 收集完性能指标(主要以毫秒为单位报告),它会通过查看指标值在其 Lighthouse 评分分布中的位置,将每个原始指标值转换为从 0 到 100 的指标分数。评分分布是从 HTTP Archive[8]上真实网站性能数据的性能指标得出的对数正态分布。

FCPinHTTP Archive

Lighthouse 评分曲线模型使用 HTTPArchive 数据来确定两个控制点,然后设置对数正态曲线的形状。HTTPArchive 数据的第 25 个百分位数变为 50 分(中值控制点),第 8 个百分位数变为 90 分(良好/绿色控制点)。在探索下面的评分曲线图时,请注意在 0.50 和 0.92 之间,度量值和分数之间存在近乎线性的关系。0.96 左右的分数是上面的“收益递减点”,曲线拉开,需要越来越多的指标改进来提高已经很高的分数。

探索 TTI 的评分曲线[9]

指标得分和性能得分根据以下范围进行着色:

0至49(红色):差

50至89(橙色):需要改进

90至100(绿色):良好

为了提供良好的用户体验,网站应该努力获得良好的分数(90-100)。

实现

static async audit(artifacts, options) {

//....
//conclusion of the lighthouse result object
const axeVersion = artifacts.Accessibility?.version;
const credits = {
'axe-core': axeVersion,
}
let categories = {};
if (config.categories) {
categories = ReportScoring.scoreAllCategories(config.categories, auditResultsById);
}
// Replace ICU message references with localized strings; save replaced paths in lhr.
i18nLhr.i18n.icuMessagePaths = format.replaceIcuMessages(i18nLhr, settings.locale);
// LHR has now been localized.
const lhr = /** @type {LH.Result} */ (i18nLhr);
if (settings.auditMode) {
const path = Runner._getDataSavePath(settings);
assetSaver.saveLhr(lhr, path);
}
// 生成报告
const report = ReportGenerator.generateReport(lhr, settings.output);
return {lhr, artifacts, report};
}

❤️ 谢谢支持

以上便是本次分享的全部内容,希望对你有所帮助^_^

喜欢的话别忘了 分享、点赞、收藏 三连哦~。

欢迎关注公众号 ELab团队收货大厂一手好文章~

智能学习前端团队自创立以来,团队专注于打破大众对教育的刻板印象,突破固有的教学思维,攻破各类教学屏障。旨在为每一位学生制定最合适的学习方案,予以因材施教,使优质教育随‘触’可达。

  • 字节跳动校/社招内推码: XW7FF6M

  • 投递链接: https://job.toutiao.com/s/MsXj3d4

可凭内推码投递 字节教育 - 智能学习前端团队相关岗位哦~

参考资料

[1]

原理结构: https://github.com/GoogleChrome/lighthouse/blob/main/docs/architecture.md

[2]

Puppeteer: https://github.com/puppeteer/puppeteer

[3]

WebSocket: https://github.com/websockets/ws

[4]

Better debugging of the Protocol: https://github.com/GoogleChrome/lighthouse/issues/184

[5]

DevTools Protocol: https://chromedevtools.github.io/devtools-protocol/

[6]

trace event: https://docs.google.com/document/d/1CvAClvFfyA5R-PhYUmn5OOQtYMH4h6I0nSsKchNAySU/preview

[7]

性能评分: https://web.dev/performance-scoring/

[8]

HTTP Archive: https://httparchive.org/reports/state-of-the-web

[9]

探索 TTI 的评分曲线: https://www.desmos.com/calculator/o98tbeyt1t

- END -

展开阅读全文

页面更新:2024-05-11

标签:分数   评分   团队   性能   指标   协议   事件   页面   方法   报告

1 2 3 4 5

上滑加载更多 ↓
推荐阅读:
友情链接:
更多:

本站资料均由网友自行发布提供,仅用于学习交流。如有版权问题,请与我联系,QQ:4156828  

© CopyRight 2020-2024 All Rights Reserved. Powered By 71396.com 闽ICP备11008920号-4
闽公网安备35020302034903号

Top