The sequence file currently stores a fixed list of metadata attributes, such as key/value class names,
compression method, etc. To make sequence file more self descriptable, it should allow to store a list of key/value pairs. One particular attribute of interest is to indicate whether the key/value classes are actually hadoop record classes,
if so, store the DDls for the records. This way, we may create tools to extract DDl from a sequence file and
then generate necessary classes. It also make it possible to provide an interpretive version of Hadoop record.
This way, even in the situation where Hadoop or the application does not have the necessary classes,
a sequence file of Hadoop records can be read and deserialized "interpretively".