When csv.DictWriter.writerow is fed a dict with extra fieldnames (and extrasaction='raise') and any of those extra fieldnames aren't strings, a TypeError-exception is thrown. To fix the issue; in csv.py, edit the line:
raise ValueError("dict contains fields not in fieldnames: " + ", ".join(wrong_fields))

to:
raise ValueError("dict contains fields not in fieldnames: " + ", ".join(repr(wrong_field) for wrong_field in wrong_fields))

Attached is a patch that fixes the problem (works in both 2.6 and 2.7, I haven't tried anything else).

Here is a simple test to demonstrate the problem:

import cStringIO, csv

sio=cStringIO.StringIO()
writer=csv.DictWriter(sio, ["foo", "bar"])
writer.writerow({1:"hello", 2:"world"})

I would argue that the TypeError is correct (field names must be strings), even though the way it is generated is a bit unorthodox :)

Let's see what others think.

Rereading my post I disagree with myself.  ValueError is probably better in this context (the difference between ValueError and TypeError is a bit grey, and Python is not necessarily completely consistent about it.)

If non-string field names aren't allowed then shouldn't they be caught at an earlier stage, rather then when the user feeds writerow a dict with an unexpected key?

But why should the field names have to be strings in the first place? Everything else is passed through str before being written anyway...

> But why should the field names have to be strings in the first place? Everything else is passed through str before being written anyway...

Good point.

Here is the preliminary patch.

New changeset 6e5afeada7ca by R David Murray in branch '3.3':
#19449: Handle non-string keys when generating 'fieldnames' error.
http://hg.python.org/cpython/rev/6e5afeada7ca

New changeset ee2c80eeca2a by R David Murray in branch 'default':
Merge: #19449: Handle non-string keys when generating 'fieldnames' error.
http://hg.python.org/cpython/rev/ee2c80eeca2a

New changeset e52d7b173ab5 by R David Murray in branch '2.7':
#19449: Handle non-string keys when generating 'fieldnames' error.
http://hg.python.org/cpython/rev/e52d7b173ab5

Thanks, Tomas and Vajrasky.  I tweaked the patch slightly: Thomas's fix was better, since it doesn't incur the overhead of the repr unless an an error is detected (a micro-optimization, true, but an easy one).  I also chose to only check that 'fieldnames' is mentioned in the non-variable exception text; exactly how the rest of the message text is worded is not part of the API.