This needs discussion, but one of the obvious possible changes for 2.3 can be to make it a Long (we cannot do it for 2.2 though). Also, I am not very comfortable with the LCG generator in ThreadLocalRandom -- unfortunately adding a new PRNG requires extensive testing (although personally I would like to have one trustworthy PRNG in future versions, even if not for this ticket).

Proposal: generate UIDs sequentially in the parent, randomize the starting point when the parent creates its first child.