Mixing Decimal and float

[B.V.]

On Jun 2, 12:22 pm, Mark Dickinson <dicki... at gmail.com> wrote:
> On Jun 2, 9:24 am, "B.V." <bv.try... at gmail.com> wrote:
>
> > Hi,
>
> > In order to solve some issues due to operations between Decimal and
> > float, we wanted to implement a class that inherits from both float
> > and Decimal.
>
> > Typically, we wrote:
>
> > class Float(Decimal, float):
>
> Can you explain exactly what issues you want to solve, and how you
> want your Float class to behave?  Do I understand correctly that you
> want your Float class to be able to represent both floats and
> Decimals?

Let me give you the whole story. We work on Tryton, an client/server
application framework written in Python (http://www.tryton.org).
The framework defines several types of fields within its own ORM
(http://doc.tryton.org/1.6/trytond/doc/ref/models/fields.html#ref-
models-fields); among those types, there's a fields.Float type -- not
to be confused with the class Float we are talking about -- (with
underlying python type float) and fields.Numeric (with underlying
python type Decimal).
fields.Numeric(Decimal) where implemented at the beginning of the fork
(Tryton is a fork of OpenERP, formerly known as TinyERP), because the
use of floats in OpenERP leads many problems in module handling
financial data.

The client is written in pygtk. The client connects the server through
a specific (but simple) protocol called pysocket (roughly pickled data
over sockets).
In an application, you may define objects with both Numeric or Float
attributes, and when you need to make them interact, you have to cast.
And everything is fine.

But trying to be open to other languages, the server implements also
an XMLRPC interface (and also a JSONRPC-like interface). That's the
key point: Decimal is python specific. So in an application, you can't
rely on the value received from a client, because depending on the
protocol, the type of the value is different.
So the idea was to create a class that can behave like a Decimal or a
float depending on the context, and set
xmlrpclib.Unmarshaller.dispatch["double"] to a function that return a
Float instance.

A contributor filed an issue on the bug tracker (https://
bugs.tryton.org/roundup/issue1575) and because he's a nice guy (ok
it's a friend of mine), he made a patch proposal (http://
codereview.appspot.com/1387041). The end of the story is in the
comments of the proposal.

>
> > But we also need to do:
> > isinstance(Float('1'), float) == True
> > isinstance(Float('1'), Decimal) == True
>
> Can you explain why you need this?

It's a requirement of the project leader.

>
> Should isinstance(Float('1.1'), float) and isinstance(Float('1.1'),
> Decimal) also both be true, or would only one of those be true?  (And
> by the way, what value would Float('1.1') have?  float('1.1') and
> Decimal('1.1') are different values.)

I think they both should be True, for '1', '1.1', '0', '0.1', ...
For the value, I would say that it depends of the definition of the
field (fields.Float or fields.Numeric).


>
> I don't think your approach can succeed;  I'd suggest just subclassing
> 'object' and abandoning the 'isinstance' requirements.  Or perhaps
> creating a subclass of Decimal that interacts nicely with floats.  You
> might also want to investigate the numbers ABC, though that's new in
> Python 2.6.

First, Float implementation was a subclass of Decimal that works with
floats, and solves many (maybe all) problems. But as you may read in
the comments of the patch proposal, it seems to be not enough.

B.