Technical detail · IoT Architecture Anti-Patterns: Diagnosis
Under the Hood: why anti-patterns survive demos
2.6 Why Anti-Patterns Survive Early Demos
Architecture anti-patterns pass early tests when scale, outages, many device software versions, and operational handoff are absent. Ten sensors can hide cloud dependence; one dashboard can hide database coupling; a short indoor demo can hide a relay energy hotspot.
Backhaul carries traffic from a local network to a wider network or cloud. Received signal strength indicator (RSSI) measures received radio signal strength; investigate it alongside retries and relay load. A p95 latency is the time within which 95 percent of measured requests finish, revealing slow cases an average can hide.
The deeper mechanism is coupling. A cloud-only design couples device behavior to backhaul availability and cloud latency. Direct database access couples applications to table names, column units, retention policy, and storage engine choices. Energy-hotspot routing couples network lifetime to a few relay nodes. Each coupling may be invisible in a pilot because the dataset is small, the field network is clean, and the same engineer owns device code, backend, and dashboard.
Use the model to make coupling measurable. Count how many application files know vendor table names. Count how many commands require cloud round trips before a local actuator (a part that moves or switches something, such as a valve) can fail safe. Count how many packets a relay forwards compared with the average node. Track p95 end-to-end latency, broker queue depth, the rate at which timestamped readings are saved, gateway disk retention, and reconnect behavior after a WAN outage. These signals tell the team whether the system needs edge filtering, a protocol bridge that translates between device communication rules, a queue, an API abstraction, another gateway, or simply a tighter operational limit.
The model should also drive the launch test. A mature design can explain which layer owns calibration, identity, buffering, normalization, command state, dashboard queries, and support escalation. It can survive adding a new firmware version, a new sensor vendor, a temporary cloud outage, and a traffic burst without every layer learning every other layer’s internals. If the answer depends on one layer silently absorbing every responsibility, the architecture needs a boundary before launch.