Spark - Different ways of reading a subset of columns from a parquet will result in different input and shuffle size

Question

Can someone explain to me why that reading the same subset of columns from parquet using different method could result in different Input and shuffle size ( when shuffled )?