At scale, the first problem you will generally hit is that each subscriber needs to hold open a connection to the database, chewing precious database resources

At scale, the first problem you will generally hit is that each subscriber needs to hold open a connection to the database, chewing precious database resources (1 process, a little RAM, maintenance overhead). Works great on smaller systems though. Hard to state where the cutoff is given it is a function of hardware capacity.