Three‑Year Full‑Chain Performance Optimization at Gaode Map: Strategies, Practices, and Results

Since 2019, Gaode Map has continuously optimized the performance of its entire core chain for three years, ultimately halving the overall latency and significantly improving user experience. This article summarizes the thinking, methodology, and practical experience of the optimization process.

Overall Approach

The approach consists of three parts: identifying performance bottlenecks, reverse‑order targeted problem solving, and forward‑order long‑term control.

1. Identify Performance Bottlenecks

• Define a scientific evaluation metric. The first‑screen load time is used as the primary statistic because it heavily influences user perception. Different device tiers are considered to ensure coverage across the user base.

• Statistical Standards

Business perspective: different pages have different first‑screen definitions.

Product perspective: prioritize high‑frequency features.

R&D perspective: use log points to anchor the start and end of the first screen.

R&D‑Test alignment: establish a unified quantitative language.

• Device Standards

Classify devices into high, medium, and low tiers based on benchmark scores.

Select representative devices for each tier according to market share and lab availability.

• Determine Optimization Items

Given the massive historical code base, manual analysis is infeasible. Tools and methodologies are required to quickly locate time‑consuming sections.

2. Reverse‑Order Targeted Problem Solving

Performance issues often involve multiple product‑research‑test teams. The process starts from the problem, works backwards, forms a专项 (special project) to gather resources, set clear goals, and achieve rapid results, thereby boosting team confidence.

Key steps include:

Collect and analyze data from low‑end devices where latency is most noticeable.

Prioritize business scenarios (e.g., travel, search) based on user behavior and BI data.

Map the full‑chain flow, insert precise log points, and extract real‑world first‑screen timings.

Special projects are launched with iterative versions, gradually reducing manpower as tools and experience mature.

3. Forward‑Order Long‑Term Control

After solving the immediate issues, a continuous control mechanism prevents regression and sustains performance gains. This includes standards, processes, automation platforms, and tools.

Optimization Solutions

Business Adaptive Resource Scheduling

A resource‑scheduling framework automatically senses the runtime environment and makes decisions (e.g., downgrade low‑priority features on low‑end devices, pre‑process user‑prefixed actions). The framework monitors execution results and feeds back for further tuning.

Engine Acceleration

Map engine: batch rendering, frame‑rate scheduling, message dispatch.

Cross‑platform engine: thread priority boosting, context pre‑loading, framework reuse, require‑module reuse.

H5 Container Enhancements

Offline package acceleration: build, publish, and manage offline bundles to achieve “second‑open” speed.

Container pre‑creation: pre‑create WebView instances during idle periods to reduce first‑time launch cost.

High‑Performance Components

Thread pool with five priority queues (high, next‑high, normal, background, main‑idle) to allocate CPU resources efficiently.

Network library: fine‑grained request scheduling, parallel pre‑processing, DNS pre‑fetch, HTTP/2 keep‑alive, adaptive compression, and protobuf parsing for heavy responses.

Platform & Automation

Gaode’s Titan CI platform and ATap automated testing provide end‑to‑end pipelines for build, performance testing, issue tracking, and rapid resolution. Features include scheduled builds, performance‑package generation, automated test triggers, integration checkpoints, and performance dashboards.

Summary

The three‑year effort dramatically improved the core‑chain performance of Gaode Map. Tactics such as专项 (special projects) + technical沉淀 (knowledge accumulation) + long‑term control ensured sustainable gains. Strategically, the team shifted from solving problems solely with manpower to leveraging architecture and tools, moving toward automated, self‑healing performance management.