Systematic Performance Optimization of JD Shopping Mini Program for Billion‑User Scale

The JD shopping mini program, serving hundreds of millions of users, suffered from slow startup performance, with an 86% launch success rate and significant user loss. To retain users and support growth, the technical team embarked on a systematic performance optimization project.

Performance Metrics and Measurement – The team defined key stages such as loadPackage , First Paint (FP) , First Contentful Paint (FCP) , First Meaningful Paint (FMP) , and Largest Contentful Paint (LCP) . Monitoring fields like _perf_downloadPackage , _perf_evaluateScript , _perf_firstRender , _perf_route , _perf_firstContentfulPaint , _perf_appLaunch , and _perf_largestContentfulPaint were collected via the Zhulong monitoring platform to compute a User Experience Index (UEI).

Optimization Measures

1. Code‑Package Size Reduction – Implemented independent sub‑packages, async sub‑packages, and pre‑download of sub‑packages to shrink the main bundle and reduce download and injection time. For example, the JD Express homepage became an independent sub‑package, cutting its startup time by 420 ms (14.5% reduction).

2. Code Injection Optimization – Enabled lazy loading of required components via configuration:

{
  "lazyCodeLoading": "requiredComponents"
}

Selective injection prevents unused components from loading, and synchronous API calls (e.g., getSystemInfoSync , getStorageSync ) were replaced with asynchronous alternatives.

3. First‑Screen Rendering Optimization – Removed unused custom components, trimmed first‑screen data, deferred off‑screen component loading, and employed data pre‑fetching ( wx.getBackgroundFetchData ) and local caching ( wx.setStorage / wx.getStorage ) to render meaningful content faster.

4. Release Frequency Management – Planned version releases to avoid excessive updates that increase startup latency, and enabled “prefer local version” to let the client use cached code when possible.

Results and Outlook – After applying these strategies, the mini program’s cold‑start time dropped from ~4800 ms to below 2500 ms, and the launch success rate rose above 90%, saving millions of users daily. Future work will continue to refine first‑render performance, explore new rendering engines (e.g., Skyline), and further streamline data payloads.