Using a python module that expected me to pass a dictionary as a
parameter to a function, I found that I was getting unexpected results
when using a class which inherits from types.DictType.

The module was passing the instance I had supplied as a parameter to
update() on a newly created dictionary. I wanted my class to present its
methods as items of the dictionary it was pretending to be. Even though
I could see that __getattribute__ had been called to get 'keys', keys()
was not called and I ended up with an empty dictionary.

Trying to find the cause of the problem, I downloaded the python source
package and found that in PyDict_Merge (in Objects/dictobject.c, line
1359), there is an optimisation that is run if PyDict_Check determines
that the parameter we passed is itself a dictionary. The optimisation
takes advantage of the fact that we know how data is stored in the
dictionary and accesses whatever it needs directly.

I had overridden keys() to provide the result I wanted but this
optimisation prevented the code from ever being run, leading to code
that seems like it should work failing with no apparent cause. The only
way I could find around this was to raise AttributeError in
__getattribute__ when getting 'keys', causing dict_update_common to call
PyDict_MergeFromSeq2 in place of PyDict_Merge.

Should the call to PyDict_Check in PyDict_Merge not be replaced with
PyDict_CheckExact ? This would keep the optimisation in place in most
cases, where we know that it is safe, still allowing keys() to be
overridden as you expect.

Will look back in my notes. I believe this was previously discussed
and rejected.

This is duplicate of http://bugs.python.org/issue1615701 . The
check_exact solution was attempted once but it broke code in a subtle
and hard to find way so it had to be reverted. For your use case, try
UserDict instead of subclassing dict.