How Huawei’s YuanRong Is Redefining General-Purpose Serverless Computing

Huawei Cloud recently announced a breakthrough: the YuanRong general‑purpose serverless platform, presented at ACM SIGCOMM 2024. The paper "YuanRong: A Production General‑purpose Serverless System for Distributed Applications in the Cloud" describes the first industry‑wide serverless system that offers a universal function programming model, high scalability, high performance, and efficient backend integration.

01 Serverless from “dedicated” to “general”

Current serverless products are mainly event‑driven and struggle with complex stateful micro‑services, big data, HPC, and AIGC workloads. Four core technical challenges are identified:

Inefficient function coordination: functions cannot address each other directly, requiring gateway hops and lacking shared memory, which harms low‑latency micro‑service and HPC scenarios.

Uncertain cold‑start latency: large container images and complex initialization cause minute‑level cold starts, unsuitable for elastic demand.

Externalized state performance impact: persisting state to remote storage (e.g., OBS) adds hundreds of milliseconds latency and limited bandwidth, hindering high‑throughput data flows.

Complex interaction between user functions and backend services: stateful services such as JDBC connections are hard to share across function instances, and distributed transactions become problematic.

02 YuanRong’s pioneering general‑purpose serverless platform

YuanRong introduces several key innovations to address the challenges above:

Scalable function system enabling massive function scheduling, sub‑millisecond inter‑function calls, and ultra‑fast cold starts.

Multi‑semantic data system providing distributed shared‑memory objects and streaming data objects, supporting efficient data flow.

Portable Bridge system offering standardized abstract interfaces for events and backend services, decoupling architecture while supporting connection reuse and shared transactions.

Cloud‑native runtime interface with language‑specific runtimes, delivering domain‑specific programming models for web services, big data, AI training/inference, and HPC.

03 General‑purpose Serverless use cases

Case 1: Global automotive leader builds a million‑scale IoV platform

The automotive IoV platform must handle minute‑level vehicle data reporting, 100 TB daily increments, and peak QPS over 30 k, while supporting over 10 PB of stored data. By adopting Huawei Cloud FunctionGraph (powered by YuanRong) together with APIG, DIS, and EG, the solution achieves on‑demand elasticity, pay‑per‑request billing, and zero‑ops management.

After migration, the platform reaches 5 000+ function instances scaling within minutes, reduces end‑to‑end latency from minutes to seconds (≈20× speedup), and improves resource utilization by 50%.

Key technologies include:

Process‑level snapshot fast cold start: functions are frozen into snapshots and restored instantly, cutting cold‑start time by over 90%.

Hierarchical scheduling architecture: enables parallel scheduling of massive function instances with data locality awareness, supporting massive vehicle connectivity.

Case 2: Huawei MetaERP full serverless upgrade

MetaERP is a SaaS system covering manufacturing, supply, procurement, and finance, with complex micro‑service, function, and big‑data workloads. The legacy architecture suffers from high development and resource costs.

By refactoring with YuanRong, MetaERP gains:

Asset accounting (MFA) workload sees cold‑start reduced to 5 seconds and elastic scaling 20×, cutting monthly resource consumption by 70%.

Sales order workload achieves sub‑millisecond inter‑function calls via YuanRong’s fast inter‑function communication, and resolves distributed transaction consistency through a Service Bridge that aggregates requests to a single Bridge instance.

04 Summary and Outlook

Huawei Cloud continues to build a full‑domain serverless ecosystem, launching products such as FunctionGraph, CCE Autopilot, CAE, DWS, and EventGrid, which collectively support the commercialisation of serverless. Looking ahead, YuanRong will focus on AI‑driven workloads, tackling fast elastic inference for large models, distributed KV‑cache pooling, multi‑model co‑scheduling, and ultra‑large‑scale distributed training, aiming to deliver ultra‑low‑cost, high‑performance, and highly available serverless AI solutions for millions of developers.

For more technical details, see the original ACM SIGCOMM 2024 paper: https://dl.acm.org/doi/10.1145/3651890.3672216