How NextRPC’s Multi‑Stage Payload Boosted Alibaba’s Double‑11 Conversion by 25%

Overview

NextRPC is an innovative RPC request pattern that returns payloads in multiple segments, similar to a multi‑stage rocket, allowing different network channels to request and respond with segmented data, ultimately delivering complex business scenarios.

Business Background

In core transaction flows such as shopping cart and order processing, Alibaba sought to integrate personalized recommendation algorithms (e.g., "Buy One More") to increase UV conversion rates across single‑item, multi‑SKU, and cross‑store scenarios.

Problems and Challenges

User‑experience conflict: personalized recommendation services have high response times that do not meet the strict latency requirements of core transaction paths.

Service‑quality conflict: introducing recommendation logic increases downstream system complexity and threatens stability; the recommendation path tolerates occasional failures, whereas the core transaction path requires deterministic success.

Resource conflict: differing deployment structures for transaction and recommendation services lead to heterogeneous machine capacity and distribution.

These issues raise the technical challenge of supporting multiple business support systems with differing requirements for user experience, service quality, and resource usage within a single request.

Technical Selection

1. RPC Model Analysis

Five common RPC models were compared, highlighting that traditional RPC returns a single response, whereas NextRPC supports streaming multi‑segment responses via a "Next → Next → Next" pattern.

2. Request Processing Model Analysis

Two processing models were examined:

Serial processing: overall latency is bounded by the deepest data dependency.

Parallel processing: concurrency (n) reduces latency to the maximum of the parallel branches, i.e., max(RTₙ).

3. NextRPC Multi‑Segment Return Mode

NextRPC combines asynchronous data‑stream push RPC with parallel processing, achieving:

Decoupling of core and non‑core business logic: core data is synchronously responded, while non‑core data is pushed asynchronously.

Separation of transaction and recommendation logic, enabling cross‑application calls and alleviating resource conflicts.

Parallelizable business logic to reduce serial wait times, improving user experience.

NextRPC Architecture

Client Architecture

Network layer abstracts Mtop/Accs channels, making them transparent to business logic.

Data rectification layer controls the sequencing of multiple responses within a single request, including asynchronous sub‑responses.

Data orchestration layer merges asynchronous data from multiple sub‑requests.

Server Architecture

The server processes the main request, emits sub‑requests via a message middleware (AttachedRequestEmitter), and after completing the main business logic, sends sub‑responses back to the client. Sub‑requests are consumed on separate topics or tags to isolate workloads, and the SDK provides metadata for handling expiration, duplication, and idempotency.

Technical Metrics

Main request success rate: proportion of main requests processed successfully.

Sub‑request success rate: proportion of sub‑requests processed successfully.

Sub‑response arrival rate: proportion of sub‑responses that reach the client after a successful main request, measurable at 1 s, 3 s, and 5 s thresholds.

Conclusion and Outlook

During Alibaba’s 2021 Double‑11 promotion, NextRPC was deployed in key transaction scenarios, delivering over a 5% increase in overall UV conversion and more than a 25% boost in optimal‑product selection cases. The team will continue to explore NextRPC for additional business scenarios, leveraging its ability to balance transaction certainty with enhanced user experience.