When dealing with data skew in Spark joins between a large fact table (where one join key has significantly more records than others) and a smaller dimension table, which method is most effective to reduce performance issues caused by shuffle skew?