If you use Python, I guess it might be because bytes in Python are in a range [0,255], whereas in Scala/Java [-128,127]. So if you use Spark built-in function to convert column type, the result might ...