DeWu Full-Chain Load Testing Platform: Design and Implementation

Before production load testing, DeWu used 1:1 replica environments, which were costly and not fully realistic. To reduce cost while maintaining high fidelity, a full‑chain production environment load testing approach was adopted.

The core challenge is data isolation: test traffic must not pollute live data. Using the Fusion scaffolding middleware, markers are propagated through RPC, Redis, DB, MQ, and threads; test data is written to shadow databases, enabling full‑chain testing at the middleware layer.

After years of using an open‑source platform, DeWu built its own platform (a customized JMeter) supporting protocols such as Dubbo, HTTP, gRPC, WebSocket, Java, and modes like fixed QPS/TPS or thread‑count based testing. The platform automatically generates comprehensive reports and runs on containerized load‑generator machines, lowering cost and improving stability.

Supporting fixed QPS mode allows precise control of traffic according to domain‑specific QPS/TPS targets, reducing preparation time and preventing accidental overload of production services.

Platform Features

Full‑chain high‑concurrency testing

Multi‑protocol support: HTTP, Dubbo, WebSocket, gRPC, JDBC, Java

Concurrent and throughput testing modes

Internal and external network testing

Online script configuration linkage

Dynamic resource pool with on‑demand container allocation

Master‑less load generation to eliminate bottlenecks

Self‑monitoring of load generators

Automatic report generation with QPS & response‑time curves

Scheduled task automation

Single‑request debugging

The architecture replaces the master‑slave JMeter cluster with a decentralized design where each node acts as an independent master, improving scalability, availability, and cost.

Testing Process

1. Pre‑test: resource estimation, JMX script development, parameter upload, QPS target setting.

2. During test: the control UI injects a BackendListener to InfluxDB; Grafana visualizes metrics such as total requests, errors, QPS/TPS, per‑interface and per‑generator stats, average and 95th‑percentile response times.

3. Post‑test: after completion, the service aggregates InfluxDB data, stores summary in MySQL, and saves raw series metadata to HDFS.

Test results show average CPU usage of 30% and memory usage of 50% on load generators, with healthy InfluxDB performance.

Future Outlook

Data sanitization and amplification for test data generation

Online adjustment of throughput

GUI‑driven JMX component editing to lower the learning curve

Support for additional protocols such as RocketMQ

Automatic pass‑rate analysis and self‑circuit‑breaker

Automated test plan generation from traffic recording

The DeWu load testing platform continues to evolve, providing a reference for building robust performance testing solutions.