openerp-expert-framework team mailing list archive

[Raphaël Valyi <rvalyi@xxxxxxxxx>]

Hello,

we while ago we debated here the usage of Floats vs Decimals in Python.
with OpenERP Floats, 0.3 means 0.29999999999999999 in Python for instance so
there is very often a 10^-17 error when loading exact data from PostgreSQL
for instance.
The whole debate was that if we pay attention in the dev, that 10^⁻17 error
wouldn't contaminate the 10^⁻2 required accuracy of the legal accounting (or
even possibly 10^-3).
Well I tend to agree with that in general. Tryton went for Decimals instead,
claiming it makes it easier to code, at the cost of some extra overhead of
course.

However, Sharoon Thomas (Openlabs) just reports us that he his watching the
credit and debit of some specific root account and
get a 0.04 delta between credit and debit: http://yfrog.com/f/afxxbp/
: 1033970.79
vs 1033970.83

Database numbers are exact on contrary:
 debit    |   credit
------------+------------
1033970.83 | 1033970.83

Sharoon says there are around 14 000 (~10^4) moves inside that account.
So we can at least expect a contamination to up to 10^-17*10^4=10^-13 for
credit and also for debit.

The delta he finds is even a bitter higher however:
>>> debit, credit = sum([x['debit'] for x in moves]), sum([x['credit'] for x
in moves])
>>> debit
1033970.8299999937
>>> credit
1033970.8299999965

that's a 10^-9 error, I'm not sure exactly why it's more than our 10^-13 we
expected by the way, any explanation?
It looks like the error propagated the squared we expected.

Now the question is how the recursivity of the __compute method in
account/account.py can amplify this into a 10^-2 error.
The bug is real and annoying, you can still gain your medal if you find out
how it propagates from 10^-17 to 10^-2 in the chart of accounts with only 14
000 moves.
I think Sharoon is investigating on his side meanwhile...

More soon hopefully...



Regards,


Raphaël Valyi
Founder and ERP Consultant
+55 21 3010 9965
http://www.akretion.com

<http://www.akretion.com.br/>