Dart Backend Practices at Xianyu: Glue Layer Development

This article describes how the Xianyu tech team adopted Dart as the language for the backend “glue layer” that coordinates between frontend clients and Java‑based domain services.

The team chose Dart because its syntax is familiar to Java/Python/JS developers, it offers strong asynchronous support (Future, async/await), and it enables a unified language across frontend (Flutter) and backend, allowing frontend engineers to write backend code.

Dart’s asynchronous model lets IO‑heavy glue‑layer code be written in a near‑synchronous style while achieving the performance benefits of true async, reducing overall response time by waiting for multiple IO calls in parallel.

To improve development efficiency, the team built an isolated development environment using Dart isolates, giving each engineer a unique, hot‑replaceable sandbox that deploys changes in seconds instead of minutes.

The architecture separates stable domain services (implemented in Java) from the fast‑changing glue layer (Dart). The glue layer fetches data from Java services via HSF, processes and assembles it, and exposes the result as MTOP HTTP endpoints for the client.

Sample code shows a Java synchronous method versus a Dart async version that uses Future.wait to parallelize two service calls, illustrating the clarity and performance gain.

ItemDetailDO queryItemDetail(Long itemId) {
    ItemDetailDO data = new ItemDetailDO();
    data.setBrowseCount(IdleItemBrowseService.count(itemId));
    data.setFavorCount(IdleItemFavorService.count(itemId));
    return data;
}

Future<ItemDetailDO> queryItemDetail(int itemId) async {
  var data = ItemDetailDO();
  await Future.wait([
    IdleItemBrowseService.count(itemId).then((count) => data.browseCount = count),
    IdleItemFavorService.count(itemId).then((count) => data.favorCount = count),
  ]);
  return data;
}

In production, the Dart‑based service handles up to 400 QPS with a 99.5 % success rate, and latency matches the Java counterpart, while code changes go from concept to deployment in under two minutes.