One App, Five ROMs, Five Ways to Die: 8 Real Android Compatibility Traps from a Chinese Health Team
These failure modes cost Chinese Android apps daily users, store rejections, and hours of debugging that never reproduce on a test device. The team's concrete detection and architecture patterns—ROM detection via system properties, deferred SDK init, and a known-issues runbook—translate directly to any team serving the domestic Chinese market, where GMS is absent and vendor behavior trumps platform documentation.
A health app project in China documents eight distinct ways domestic Android ROMs break otherwise working apps. The starting point is a null pointer that crashed on five latest flagship phones after each vendor pushed a system update; the root cause turned out to be a changed ContentProvider initialization order. From there the piece expands into a systematic catalog: background processes killed with no callback on five different lifespan policies, push channels that silently drop messages without vendor SDKs, overlay permissions that pass system checks but still fail at runtime, WebView kernels that drift independently of OS version, privacy compliance failures triggered by third-party SDK initialization before user consent, five app stores with different and contradictory review requirements, broadcast interception that breaks sync features, and new cracks from foldables and Android 16.
The team's response is less about fighting vendors and more about designing for failure. They built a ROM detection layer based on system properties, a device matrix that includes HarmonyOS NEXT as a separate target, a known-issues document that turns every incident into a runnable check, and a self-check script that reports permission states at startup. Their core claim is that compatibility cost lives in discovery and memory, not in fixing: a known issue takes minutes to resolve, but finding it the first time can take days. The final stack is a set of reusable patterns—unified push adapters, a PrivacyGate for SDK initialization, a compatibility matrix, and crash attribution by ROM type and targetSdk version—that any team shipping to the Chinese market can adapt.
The same null pointer appearing on five vendors simultaneously shows that third-party SDKs assuming a standard initialization order are a systemic risk, not a one-vendor quirk.
Vendor battery management is designed to protect benchmark scores and user perception of battery life, so working against it with keep-alive tricks is a losing game; the only sustainable path is user-granted permission plus architecture that survives death.
The shift from "process alive" to "dead but recoverable" is the real lesson; the cost is not in writing recovery code, but in discovering the failure modes and remembering them, which the KNOWN_ISSUES.md makes systematic.
Treating store rejections as predictable rather than avoidable reframes the problem: material errors can be eliminated with templates and checklists, while privacy issues often hide inside third-party SDKs beyond your control.
The WebView drift—kernel version decoupled from OS version—means feature detection at runtime is the only sane policy, and core business flows should avoid H5 entirely.