When I learned for my SCJP, I did so with a book, which also
had some sample questions, and some of those really shook me,
as in some cases I thought the answer in those samples was wrong
(which it turned out to really be), and others that in my eyes
were "applied hairsplitting on natural language".

Then there was the exam, during which I felt quite relieved, since
there were no hairsplitting-type-questions at all.  My feeling was
that there were only two questions where I actually tried to guess
the answer.  In the end I had only 86%, which was still enough, of
course, but left me somewhat disappointed.
What left me feeling bad was, that I'd never ever be able
to see, where I was actually wrong.

What does this "sometimes" depend on?   Just wondering...

PS: I think I learnt most, apart from the book which taught me all
the then-new 1.5 features,  by programming some toy program (in my
case a medium-level sudoku-solver) in java, and trying to do that
as "Object-oriented" as possible :-)   It was only through this
programming that I stumbled over classes like "Scanner", which
then turned out to be useful to know ...