[Helmut10001]

All of these issues with UIDs make me believe that we should maybe transition to Probabilistic Data Structures and group users randomly together, e.g. based on HyperLogLog abstracted UIDs. Only the user device itself would have the full ID, the service would get an abstracted, probabilistic version of it, which can (and will) collide with other abstracted IDs. Thus, the service could never be 100% sure who exactly a single user is - out of a group of (e.g.) 12 people that happen to yield the same probabilistic representation.

(I know there're also many issues with this approach, so take it with a grain of salt)