How Gaode Map Achieved Zero‑On‑Site Duty with Serverless Architecture

Introduction

Gaode Map’s navigation voice prompts handle billions of requests daily, especially during holidays. System instability directly affects user safety, so the team needed a more reliable and elastic architecture.

Business Challenges

Frequent UI style changes require rapid iteration.

Increasingly complex recommendation strategies lead to tangled backend code.

Large client bundles increase bandwidth cost and hinder fast updates.

Peak traffic during morning and evening rush hours creates a high QPS/low‑QPS swing, making static capacity provisioning inefficient.

Technical Choice: Serverless

Serverless was selected for its low cost, zero‑ops model, and millisecond‑level auto‑scaling. Internal measurements showed a 38% increase in development efficiency and the ability to handle >400 k QPS for core services.

Architecture Design

The platform combines two layers of Function‑as‑a‑Service (FaaS) with containerized micro‑services:

Frontend FaaS (SFF) : Built on Node.js, moves client‑side logic to the cloud, providing a unified API layer between the app and backend services.

Business FaaS : Implemented in Java or Go, encapsulates recommendation strategies and data‑transformation logic, enabling fast strategy updates without touching the main codebase.

Both FaaS layers are deployed alongside traditional container services, leveraging the efficiency of functions and the flexibility of containers.

Implementation Details

Development is powered by the Serverless Devs toolchain, which provides:

Multi‑environment CI/CD pipelines.

Gray‑release (canary) deployment.

Built‑in observability (logs, metrics, tracing).

Active‑active deployment across multiple regions.

Typical workflow: s init to scaffold a function (≈1 minute), s deploy to publish (≈5 minutes).

Traffic control is achieved through function‑level concurrency limits and QPS throttling, allowing fine‑grained flow control.

Performance Optimizations

Cold‑start mitigation : Reserved instances are provisioned for predictable traffic; scaling can be triggered by concurrency or CPU thresholds so that new instances are ready before receiving requests.

Disaster recovery : Serverless’s rapid elastic provisioning eliminates the need for idle standby machines, enabling cost‑effective multi‑region resilience.

Runtime / Container Migration

To migrate existing C++, Go, and Rust services, custom runtimes were built that embed internal middleware, allowing legacy services to run unchanged on the Serverless platform. When internal runtimes cannot satisfy external dependencies, Alibaba Cloud’s Custom Runtime/Container feature is used, requiring only a few lines of startup script changes.

Key Metrics

In 2021 the Serverless platform served >400 k QPS for core Gaode services. The architecture supports:

Millisecond‑level auto‑scaling.

Three‑zone active‑active disaster recovery (cost of a single zone).

Cold‑start latency under 100 ms for latency‑sensitive workloads using reserved instances.

Future Outlook

Beyond online request handling, the roadmap includes applying Serverless to offline workloads such as map data processing, image cropping, and message consumption, further exploiting Serverless’s elasticity and cost benefits.