Unreachable Node
A component you cannot log into, restart or inspect, which removes every operational technique that assumes access to the running system.
Server-side operations rest on an assumption that is invisible until it is absent: you can reach the thing. You can attach a debugger, read a log, restart a process, roll back a deployment, or in the worst case terminate the instance and let a new one start.
None of it applies to a phone in someone's pocket, a sensor on a wind turbine, or a gateway in a shop with no network. The consequences run through the whole design.
Diagnosis must be designed in advance, because you cannot go and look — whatever telemetry was not built is not available later. Recovery must be autonomous, because nobody will restart it; watchdogs, self-healing and safe-mode boot are requirements rather than refinements. Updates must be reversible by the device itself, since a bad update that prevents the device from reaching the network is unrecoverable remotely. And old versions persist indefinitely, so every server-side interface must tolerate clients that will never be upgraded.
The design instinct this demands is pessimism about your own future access, which sits uncomfortably with habits formed on systems you can always reach.