> If two machines are generating ids independently and happen to choose initial random positions near each other, the probability of collision is much higher than the basic birthday bound.

But the chance of the initial random positions being near each other is very very low.

If you pick a billion random numbers in an 80 bit space, the chance you have a collision is one in a million. (2^80 / (2^30)^2)

If you pick a thousand random starting points and generate a million sequential numbers each, the chance your starting points are sufficiently close to each other to cause an overlap is one in a trillion. ((2^80 / 2^20) / (2^10)^2)

In that one in a trillion case, you'll likely end up with half a million collisions, which might matter to you. But if you care about 0 collisions versus 1+ collisions, pick the monotonic version.

Right, but the point is there's no reason to accept this limitation. Likewise why hardcode millisecond scale timestamps in a world where billions of inserts per second are practical on a single server?

Or if what you want is monotonic distributed timestamps, again, HLC is how you do that properly.

So you're embracing this weird limitations for no real benefit.

And as you can see in the rest of this comment thread, a lot of people simply do not even know this behavior and are assuming the 80 bit portion is always random. Which is my whole point about having a not really an invariant invariant just being a bad way to go fundamentally.

Edit: just to reply to the below since I can't do so directly, I understand the arithmetic here. What I'm saying is there's zero reason to choose this weird scheme vs something that's just the full birthday bound and you never think about it again.

As another comment points out: just consider neighbor guessing attacks. This 80 bit random but not random field is a foot gun to anyone that just assumes they can treat it as truly random.