targets has two major challenges with pseudo-random number generation:

1. Statistical reproducibility. Different runs of the same pipeline must give the exact same results.
2. Statistical independence. Pseudo-random numbers from different targets must be independent.

For years, targets has controlled (1) in a simplistic way: set a deterministic seed for each target, where each target seed depends on the target name and an overarching global seed:

But this approach may violate (2). Different targets have different pseudo-random number generator states, and there is nothing to prevent the PRNG sequence of one target from overlapping with that of a different target.

This problem is not unique to targets, it is a general issue with parallel computing.

C.f. wlandau/crew#113.