Who is listening to a diagnostic stream: Mob.Defect.Bus (:defect_bus) and Mob.Event.Trace (:event_trace).

The write paths read the list from :persistent_term and send/2 to each pid, with no process on the hot path. This process only changes the list, and monitors every subscriber so one that exits is pruned.

A subscriber on another node whose connection drops (:noconnection) is parked, not pruned: it leaves the published list, so no emit sends to it (a send/2 to an unconnected node dials it, which over adb cannot succeed), but its topics and meta are kept. When its node reconnects it is monitored again and republished, so a host shell that re-dials the device keeps receiving without re-subscribing; if the pid died meanwhile, the new monitor prunes it. A parked subscriber whose node stays away for config :mob, :subscriber_park_ms (default 10 minutes) is dropped.

It is separate from the stores' table owners on purpose. An owner that also did subscriber bookkeeping could lose every recorded defect to a bug in that bookkeeping. And if this process dies, the published lists and the parked subscribers survive in :persistent_term, so delivery continues, and the next instance re-monitors every pid it finds there.

A subscriber is any pid, including one on a connected node: :rpc.call(node, Mob.Defect.Bus, :subscribe, [self()]) subscribes the calling shell, where passing no pid would subscribe the short-lived process :rpc runs the call in.