TestPG Load‑Testing Platform: Precise Pressure Control Architecture and Practice

Gaode (Amap) provides a national‑level travel service platform whose stability is critical, especially during peak holidays such as Chinese New Year and National Day. To ensure service reliability for millions of users, Gaode conducts full‑link load testing before major holidays, evolving from a monthly to a daily routine.

The TestPG load‑testing platform was launched in September 2018. After its first holiday deployment in 2019, the platform has been continuously improved, now capable of completing full‑link tests for three data‑center regions nationwide within one day, and the overall testing cycle for large holidays has been reduced to three days.

Two main problems hinder efficient pressure control during full‑link testing:

Low efficiency of pressure adjustment – frequent start‑stop cycles and manual estimation of required load generators make the process time‑consuming.

Coarse‑grained control – pressure is adjusted at the level of whole machines, leading to inaccurate QPS targets.

The solution focuses on two aspects: pre‑test data (corpus) preparation and precise pressure regulation during testing. By platform‑based corpus production and fine‑grained pressure adjustment, the TestPG platform achieves accurate, real‑traffic‑like pressure models.

Technical Architecture

The platform follows a master‑slave structure: a master node (Stress Controller) schedules slave load generators. An API Gateway receives user pressure‑control commands and forwards them to the Stress Controller. Redis is used for dynamic configuration caching, delivering pressure‑control directives to each load generator either via pull or publish/subscribe.

The pressure‑control center acts as the brain, issuing commands based on control strategies and receiving feedback from the load‑generator cluster. Feedback (metrics such as QPS, latency, CPU usage) is stored in a TSDB, enabling the controller to assess whether adjustments succeeded and to decide subsequent actions.

Cluster‑Level Pressure Control

Operators can specify a target QPS at any time. If the target is lower, the platform quickly reduces the output of the cluster. If higher, additional load generators are automatically allocated from a pool to meet the target, eliminating manual estimation.

Interface‑Level Pressure Control

The platform supports dynamic adjustment of traffic share for each API interface during an ongoing test. This capability allowed more than a hundred adjustments during a single National Day test, ensuring smooth execution despite rapid traffic growth.

For fine‑grained control down to feature‑parameter level, the system leverages intelligent corpus generation to extract and distribute traffic based on interface parameters, enabling precise pressure tuning for critical business scenarios.

Summary

Through precise pressure control at both cluster and interface levels, TestPG significantly improves full‑link testing efficiency, shortens testing cycles, and brings test traffic closer to real‑world conditions. The platform also highlights the importance of realistic test corpora and anticipates future enhancements such as multiple test types (load, stress, capacity estimation) and advanced traffic models powered by machine‑learning‑driven corpus intelligence.

Future Outlook

Beyond current usage, the precise pressure‑control technology can be extended to support varied testing scenarios, including peak‑spike and pulse‑load tests, providing richer insights into system capacity and resilience.