optimizing pyspark performance with schema