I am trying to do some custom packing/unpacking and am not having much luck.

Basically, I have a CF like this:

create column family CF
  with comparator = 'CompositeType(UTF8Type, LongType, UTF8Type)'
  and key_validation_class=LongType
  and default_validation_class=UTF8Type;

And pycassa happily marshals the data like so:

And I am manually converting the column names from a tuple of (u'astring',
123, u'anotherstring') into a delimited str (e.g.
"astring,123,anotherstring") for use in a url. When storing new columns, I
have to change it back to a tuple, and have to do tuple(s.split(',')).

What I want is to have these conversion steps on the column names be
transparent. I fiddled with a couple of different techniques:

class MyType(CassandraType):

    @staticmethod
    def pack(val):
        return val.split(',')

    @staticmethod
    def unpack(val):
        return ','.join(['%s'] * len(val)) % val

This obviously doesn't work because by the time these methods are called,
the value passed in is somewhat unpacked:

'^TastringP*i!^Tanotherstring'

looks like there's some padding character in the front of the strings, ^T,
and I can unpack the number in the middle using struct.unpack('>I', 'P*i!'),
but I have no way of telling where the LongType in the middle of the
CompositeType is?

If I try inheriting from CompositeType instead of CassandraType but try the
same static methods, I get the same results. It seems like the pack and
unpack methods aren't running at the very last step, at least in the case
of CompositeTypes?

btw, I am on pycassa 1.6.0, but I tried in 1.7.2 also and got the same
results.

Thanks, hopefully there is enough info here...I tried tracing the
marshaling code for a bit but I don't know enough about Thrift or the
underlying storage protocol for Cassandra so it was difficult to follow :-\