In official documentation you can read that:
Indicates that the collection operation does not commit to
preserving the encounter order of input elements.
UNORDERED essentially means that the collector is both associative (required by the spec) and commutative (not required).
Associativity allows splitting the computation into subparts and then combining them into the full result, but requires the combining step to be strictly ordered. Examine this snippet from the docs:
A a2 = supplier.get(); accumulator.accept(a2, t1); A a3 = supplier.get(); accumulator.accept(a3, t2); R r2 = finisher.apply(combiner.apply(a2, a3)); // result with splitting
The last step,
combiner.apply(a2, a3), must be applied in exactly this order, which means that the entire computation pipeline must track the order and respect it in the end.
Another way of saying this is that the tree we get from recursive splitting must be ordered.
On the other hand, if the combining operation is commutative, we can combine any subpart with any other, in no particular order, and always obtain the same result. Clearly this leads to many optimization opportunities in both space and time dimensions.