Construction of a Full-Link Load Testing Simulation Measurement System for Didi Ride-Hailing

In order to ensure stability of Didi’s ride‑hailing services during holidays and major events, the team performs full‑link load testing and needs a way to quantify how closely the test scenarios resemble real‑world traffic. Since subjective verification is unreliable, a simulation‑degree measurement system was built to scientifically evaluate the gap between load‑test coverage and production traffic.

Since 2020 the ride‑hailing load‑test team has been constructing this measurement framework and has applied it in production. The article systematically introduces the construction process.

Background – The increasing complexity of ride‑hailing business, tidal traffic spikes during holidays, and the need for stability make full‑link load testing a core technique. It simulates massive user requests to discover potential stability risks and verify system capacity.

Challenges – Major uncertainties when simulating real scenarios include time and space randomness, order distance and driver travel time variability, order‑matching randomness, and resource contention from other business lines.

To achieve “as much load as can be withstood”, a simulation‑degree measurement system was designed with two purposes: (1) reveal current test coverage and build trust, (2) identify weak points for targeted improvement.

Measurement Process – The system consists of four stages: defining measurement requirements, constructing the measurement framework, validating measurement effectiveness, and engineering implementation.

Five measurable dimensions were finally selected: interface coverage, scenario coverage, category coverage, link coverage, and module coverage. Each dimension has concrete sub‑metrics (e.g., interface coverage rate, traffic‑vs‑target ratio, weight factors, etc.).

Measurement Methods – Data normalization (directionalization, dimensionless processing, normalization) precedes metric calculation. For interface traffic achievement, two possible variables were considered: absolute difference vs. ratio to target traffic; the ratio proved more stable.

Traffic prediction models combine historical growth, weather, and calendar effects to estimate peak load for upcoming holidays, achieving <5% error on key business indicators.

Data denoising removes outliers and filters low‑traffic items (e.g., router < 100 qps, inrouter < 500 qps) to avoid skewed scores. Black‑/white‑list mechanisms handle scenarios where traffic data is unavailable or critical low‑traffic interfaces should be retained.

Scoring formulas incorporate weight factors and thresholds to keep scores stable as new samples appear.

Model Construction – Detailed formulas for interface, module, and link coverage are presented, combining traffic ratios, weight factors, and statistical fitting (e.g., linear regression, R², MSE variants).

Practical results show improved coverage visualization, accurate traffic‑vs‑prediction curves, and clear identification of weak links.

Overall Effect – The measurement system quantifies simulation degree across five dimensions, providing objective data for model calibration, targeted coverage improvement, and risk mitigation.

Conclusion & Outlook – While the current five dimensions cover many aspects, further dimensions (IDC resources, network bandwidth, etc.) will be explored to enhance precision. The system will evolve into a platform capability that supports capacity planning, intelligent load testing, anomaly detection, and broader business empowerment.

Finally, the article acknowledges contributions from the load‑test platform, data, business, and infrastructure teams, and invites collaboration for continued improvement.