- The order of sample stream (and thus the order in which loads are
triggered) currently depends on a Set and thus on the hash codes
of the objects that change with every run. Changing to a List solves
this problem.
- The FakeAdaptiveDataSet directly created a static Random (with random
seed) to compute the variation of chunk sizes. Changing this to an
injected Random object that can always be initialized with the same
seed also removed this randomness from the tests.
PiperOrigin-RevId: 353878661