Each pusher has its own loop which runs for as long as it has work to do. This
should run in its own background thread with its own logcontext, as other
similar loops elsewhere in the system do - which means that CPU usage is
consistently attributed to that loop, rather than to whatever request happened
to start the loop.