Adopting ServiceMesh (AliMesh) in Alibaba's Xianyu: Architecture, Performance Evaluation, and Lessons

Background : In Alibaba's server‑side development, Java dominates. Teams using heterogeneous languages need a way to call existing Java services and integrate with the group’s middleware.

Current situation : Before ServiceMesh, Dart services called middleware client SO libraries via C/C++ extensions (JNI‑like). This caused high operational coupling, increased complexity, high integration cost, and delayed feature delivery.

ServiceMesh concept : ServiceMesh extracts the microservice client core into an independent Sidecar proxy deployed per host. Business processes communicate with the outside world through the Sidecar.

Communication patterns : Two common approaches – iptables traffic redirection and a lightweight Mesh client that directly connects to the Sidecar. AliMesh adopts the lightweight client.

Benefits : Business code becomes simpler, troubleshooting clearer, and the heavy lifting of the microservice framework is handled by a dedicated Mesh team, allowing independent upgrades without affecting business services.

AliMesh HTTP scheme : Provides HTTP and HSF access; HTTP further splits into HTTP1.x and HTTP2.0. In the data plane, the business process talks to the Sidecar via HTTP/JSON, and the Sidecar calls Java HSF services. The control plane syncs service lists via ISTIO pilot. Performance tests showed CPU usage roughly doubled and RPC latency increased ~2 ms for HTTP1.x, which did not meet online requirements.

HTTP2.0 improvement : Connection multiplexing, header compression, and other HTTP2.0 features reduced CPU increase to <20 % and added only ~1 ms latency per RPC.

HSF private‑protocol scheme : Uses HSF 2.0 with Hessian 1.0 encoding. The Sidecar syncs service lists directly with Configsvr via differential updates, reducing control‑plane CPU load. Optimizations kept Mesh‑induced CPU growth under 20 % and added less than 1 ms to RPC latency.

Deployment in Xianyu : The Dart + AliMesh solution has run stably for over eight months, serving detail pages, recommendation feeds, and housing homepages. Monitoring shows only a negligible impact on response time after mesh adoption.

Conclusion : ServiceMesh decouples service‑to‑service communication from business logic, slimming down applications and allowing heterogeneous language teams to integrate with existing systems at low cost. The performance overhead is modest, and the overall benefits outweigh the drawbacks.