When using cPickle to pickle / unpickle an object instance whose
__dict__ contains a dictionary of NumPy Arrays (on a windows32 system),
some of the array elements have the wrong type raising a ValueError:
could not convert string to float.  

On UNIX platform this error does not occur, and the data is read out in
the correct type every time.  
Forcing the caller to use pickle.py module instead removed the issue.

Statements about the imprecision of cPickle (such as issue1536, 655802),
or its deprecaition (now I can't find where that was mentioned), would
assist.  By contrast the current state of the documentation implies that
cPickle is better overall, and thus should be used preferentially.

Could you provide a test case? The behaviour you are describing sounds
like a bug in cPickle.

I have now pinpointed the error to a list of infinities (see attached).
When using pickle.py to read the cPickle'd data we get a different, and
more, informative error:
ValueError: invalid literal for float(): 1.#INF

Thanks for the test case. I will take a look.

Could you give me the output of this?

  import cPickle
  print repr(cPickle.dumps([float('+inf'), float('-inf'), float('nan')]))
  print [float('+inf'), float('-inf'), float('nan')]

By the way, are you sure this bug occurs on Python 2.6? Python 2.6 uses
a platform-independent float to string converter (i.e.,
PyOS_double_to_string) which shouldn't output stuff like "1.#INF"

Also, can you verify that the bug does not occur with pickle protocol 1
and over?

Your test prints:
'(1p1\nF1.#INF\naF-1.#INF\naF-1.IND\na.'
[inf, -inf, nan]

My installation is Python 2.6.2 as currently distributed.

Specifying protocol 1 or 2 does circumvent the error.  
Thank you.

I think this can be closed.  It should no longer be a problem in Python 2.7 or Python 3.x, and there's a workaround (use protocol 1 or 2) for Python 2.6.

In theory, it *could* still be fixed for Python 2.6.6, but changing the pickle output in a bugfix release seems like it might be dangerous.

However, here's a patch.  I haven't tested it on Windows.

It is too late for 2.6.6 now that it is released.