[vecter]

Wouldn't that unevenness only affect 2^31 - 2 and 2^31 - 1, so a negligible fraction of the integers? Was that tiny discrepancy enough to make your calculations?

In other words, what do you mean that it was done by doing mod 3 of a signed int32? If it was a monotonically increasing or random int32, I don't see how that unevenness would manifest in a meaningful way.

[madrox]

In another subthread, we realized my memory was wrong and we were measuring millisecond collisions. The serving ID imbalance was a side-effect. Also, it might've been an int16 I was thinking of but turns out the whole thing was shadows on cave walls.