NextRPC: Multi‑Stage RPC Model for Scalable Transaction and Recommendation Services

NextRPC is an innovative RPC request model that returns payloads in multiple stages, allowing different network channels to deliver partial responses sequentially. The “Next” in its name reflects its role as the next‑generation RPC beyond traditional single‑response patterns.

Deployed in Taobao’s transaction scenarios (e.g., order‑exchange and “buy‑one‑more” recommendation), NextRPC has demonstrated over 5% UV conversion improvement during major sales events, with up to 25% uplift in specific recommendation flows.

Business challenges addressed: conflicting user‑experience latency requirements, service‑quality stability for core transaction paths, and heterogeneous machine‑resource allocations when integrating personalized recommendation services.

Technical choices: comparison of five RPC models, analysis of serial vs. parallel request processing, and adoption of a hybrid model that combines asynchronous data‑stream pushing with parallel execution.

Client architecture: network layer abstracts Mtop/Accs channels, data‑rectification layer orders multi‑response streams, and data‑orchestration layer merges asynchronous results.

Server architecture: main request handling triggers auxiliary sub‑requests via a message broker; sub‑responses are processed independently (isolated by topic/tag and thread pools) and pushed back to the client.

Technical metrics: main request success rate, sub‑request success rate, and sub‑response arrival rate (including 1s/3s/5s latency buckets) are used to evaluate service quality.

Future work includes extending NextRPC to more business scenarios, further improving transaction certainty and user experience while empowering merchants with higher operational efficiency.