Zero‑Instrumentation Interaction and Performance Monitoring for Large‑Scale Mobile Apps

Zhang Zitian, the client technology director at Qunar, shares his experience from a Ctrip technical salon where he discusses the challenges of large‑scale mobile client development, especially the difficulty of quickly locating obscure crashes and performance problems.

Typical user reports only contain limited information such as crash time, page, and reason, making traditional crash‑log analysis inefficient and often requiring many engineers to be involved late at night.

To address this, the team proposes a three‑layer solution: a no‑track interaction statistics platform, performance monitoring, and exception monitoring, all aimed at reconstructing the complete user journey without intrusive instrumentation.

The no‑track interaction platform (QAV) collects interaction logs from three independent systems—untracked interaction statistics, performance monitoring, and exception monitoring—by using a unique widget identifier based on XPath, which overcomes the limitations of view IDs and coordinate‑based methods.

Technical implementation relies on compile‑time AOP using a JavaAgent. Custom Gradle tasks (installInject and uninstallInject) attach the agent to the JVM, inject bytecode with ASM, and ensure the agent is removed after the build, providing a non‑intrusive way to capture interaction and network data.

Performance monitoring extends the same AOP technique to capture network requests. While iOS allows runtime hooking via provided APIs, Android requires compile‑time instrumentation because runtime dex manipulation is impractical.

Data aggregation is achieved by compressing and uploading logs, generating a requestId for each interaction, injecting it into subsequent network calls, and aligning all events using synchronized timestamps, thereby creating a unified timeline of user actions, network activity, and exceptions.

The result is a reliable reconstruction of user sessions that dramatically reduces debugging time, lowers communication costs, and integrates into the broader "Jindouyun" system that supports the entire app lifecycle from development to operation.