Another fun one I’ve considered is precomputing the dependent nodes we already know about rather than reconstructing each time a task is executed to see if it has additional ones.