Here's a snippet from Python 2.6:

>>> from decimal import Decimal, getcontext
>>> getcontext().prec = 3
>>> Decimal('NaN').max(Decimal('1234'))
Decimal('sNaN234')

The result here should be Decimal('1.23E+3')---the specification says that 
the result of comparing a quiet NaN with a finite value should be that 
finite value, rounded according to the context.

This also affects min, max_mag and min_mag.

The cause is that non-NaNs are incorrectly being passed to the _fix_nan 
method.  The attached patch fixes this, and adds new testcases to 
extra.decTest to stop it happening again.

It would be good to get this fix into 3.0, if possible.  I think it should 
also be backported to 2.5.3.