[Python-Dev] Decimal & returning NotImplemented (or not)

[Nick Coghlan]

A recent question on c.l.p pointed out that the 2.4 Decimal implementation 
raises TypeError directly for operator arguments it doesn't understand, instead 
of returning NotImplemented.

Obviously, this creates problems for anyone trying to define a class that 'plays 
nicely' with Decimal (but does not inherit from Decimal), since their __rop__ 
methods never get called - Decimal's TypeError gets in the way.

I had another look at the PEP and didn't see anything on this topic, and I can't 
recall any discussion on the topic either on this list, or directly with Raymond 
and Facundo (Google doesn't have anything, either).

My current thoughts run as follows:

1. The standard binary operations (that is, '/', '*', '+', '-', '%', '**', '//', 
divmod(), cmp()) should be returning NotImplemented instead of raising TypeError 
directly.

2. The method-only operations should continue to raise TypeErrors

3. Invocations via Context methods should continue to raise TypeErrors

I was going to say this couldn't be fixed for Python 2.4 because it was a 
semantic change. But I modified my local copy to work this way, and the full 
Decimal test suite passed without any issue.

For syntax-based access, the bad call still results in a TypeError from the 
PyNumber_* machinery, and there isn't anything in the docs that says anything 
about whether a bad operand in a direct call to a special method results in 
NotImplemented being returned or TypeError being raised.

So, 2 questions:

1. Should Python 2.5's C implementation of Decimal follow the standard numeric 
coercion rules as described above?
2. Is it reasonable to class this as a bugfix and fix the Python version for 
2.4.2? (2.4.1's a bit too soon for my liking)

Cheers,
Nick.

-- 
Nick Coghlan   |   ncoghlan at email.com   |   Brisbane, Australia
---------------------------------------------------------------
             http://boredomandlaziness.skystorm.net