And again, Spark's Naive Bayes throws some exception about too much features whilst scikit works