Ele.me improves performance load times with a multi-page Progressive Web App
Ele.me is the biggest food ordering and delivery company in mainland China. It serves 260 million registered users from over 200 cities all around China, and has over 1.3 million restaurant listings. With 99% of its users ordering food on mobile, Ele.me set out to improve its mobile web experience, making it faster and more reliable in flaky connections, all while relying on the core technical model of a multi-page app to accommodate their operational needs.
- Loading time decreased by 11.6% across all pre-cached pages
- Loading time decreased on average by 6.35% across all pages.
- Time-to-consistently-interactive dropped to 4.93 seconds on a 3G network on first load
After we released the ele.me PWA, our loading times have dropped significantly, transforming our mobile web experience into one of the fastest food reservation sites in China.
Spencer Yang, Product Manager of Ele.me PWA
Choosing between multi-page app and single page app #
The explosive growth of Ele.me in recent years has led to the growth of distinct business units within the company, each in charge of running its micro-service under the main https://ele.me domain. The Ele.me team concluded that the decoupling of these individual services is best served by a multi-page app (MPA) model, with each team running and maintaining its own service.
Applying PRPL to a MPA #
The PRPL pattern (Preload critical resources, Render initial route, Pre-cache remaining routes, Lazy-load remaining routes) provides web developers with a set of rails to guide the structure of a PWA, with a particular emphasis on a quick time to interactivity and maximizing caching to reduce network round trips. While PRPL has been well-tested on SPAs, it was less clear how one would actually apply it on a MPA. Ele.me decided to take on the PRPL mindset when thinking about a rebuild of their MPA as a PWA. To do that, they make sure that when a user navigates to a page, that they are preloading critical resources for that page by including
<link rel="preload"> as needed, or surfacing those scripts at a sufficiently shallow level so that the browser’s preloader can do its work without needing additional hints. They also progressively enhance their PWA by installing a service worker whenever it is supported by the browser, which they then use to fetch and pre-cache other top-level navigation routes so that the user gets a faster loading and rendering experience as they click around the PWA. Each page in a MPA is its own route, so speeding up the rendering of the initial route is equivalent to implementing best practices to optimize the critical rendering path for each route. With these changes, the overall loading time decreased on average by 6.35% across all pages.
Serving the transition skeleton screens ASAP #
Ele.me wanted to apply the idea of skeleton screens into the UX, which is a way to ensure that whenever the user taps any button or link, the page reacts as soon as possible by transitioning the user to that new page, and then loading in content to that page as the content becomes available; this is also the key to improving the perceived performance of the PWA. However, since each page in a MPA is its own initial route, each navigation requires redoing all the necessary loading, parsing, and evaluation work every single time.
To work around this, Ele.me built the skeleton screen as an actual UI component, and then used Vue.js’ Server Side Rendering stack to build and then pre-render the Vue components to strings before injecting them into HTML templates. This allows them to render the skeleton screen directly and achieve a more fluid transition when navigating between pages.
Caching shared resources with service worker #
Routes that are considered important, but not critical, are incrementally cached by the service worker at runtime instead as the user continues to navigate through the PWA. This allows Ele.me to serve the PWA to users directly from cache under all network conditions. The result: loading time decreased by 11.6% across all pre-cached pages.