The root cause for this issue is due to a combination between the multithreaded executor and the mutually exclusive callback group used for all the ROS topics.
When the executor collects all the references to the subscriptions, timers and services, it skips the mutually exclusive callback_groups which are currently locked (ie: being processed by another thread).
This cause the resulting waitset to only contain the guard pointers.
If there is no activity on those guards, the thread will wait for work forever in the get_next_executable and block all other threads.