The article refers interchangeably to tables and collections. Reading this[1], it seems they are using MongoDB as their transactional database. If they do, it would be interesting to know at what point did they decide to split the subscriptions data into its own collection. The article just states "As we added new features, this data model became problematic."
Mongo-modelling being a black art of sorts (to me, at least) I'd be more interested in what the tipping point was for them (data size and shape, usage patterns) than the relatively straightforward (conceptually at least, operatinally these things are never to be underestimated) table-doubling approach to changing a data model.