What does parallelStream() change compared with stream()?

The pipeline is the same; the execution is not. The source is split by its Spliterator, the chunks run as fork/join tasks in ForkJoinPool.commonPool(), and the calling thread joins in rather than waiting idle. The common pool has one thread fewer than there are cores, so with the caller the work uses every core.

Three consequences are what the interviewer is after:

• Order is not guaranteed for forEach. Use forEachOrdered, or better, a collector.
• Side effects break. Adding to a shared ArrayList from forEach loses elements or throws. collect(toList()) is correct in parallel because each chunk builds its own list and they are merged.
• It is not automatically faster. Splitting and merging cost something, so a few thousand cheap operations usually run slower in parallel. Measure first.

And the one that bites in production: every parallel stream in the JVM shares that one common pool, so a slow or blocking operation in one of them slows down all the others.