python / cpython

The Python programming language
https://www.python.org
Other
62.59k stars 30.03k forks source link

`_decimal` and `_pydecimal` compatibility differences #117056

Open antonok-edm opened 6 months ago

antonok-edm commented 6 months ago

Bug report

Bug description:

There are two backends for the decimal module, one from libmpdec (_decimal) and the other being a native re-implementation in Python (_pydecimal).

The signatures of methods like Decimal.__add__ take a named context argument when imported from _pydecimal, but cannot accept that argument when imported from _decimal.

mpdecimal is not bundled with Python installations on all platforms, and my understanding is that there are even more changes to how it's packaged incoming as per https://github.com/python/cpython/issues/115119. The difference between the two backends has caused cross-platform incompatibility for at least one project, and I'd argue it's very difficult for anyone who encounters this problem to understand the issue because the error messages mention nothing about an external dependency being required.

Would it be possible to remove the context argument from the _pydecimal backend, or to add it to the _decimal backend even if it's unused, to resolve the incompatibility?

CPython versions tested on:

3.11

Operating systems tested on:

Linux

serhiy-storchaka commented 6 months ago

It is other way: the Python implementation accepts and passes the context argument to some special methods, while the C implementation has standard signatures.

The implementation is not designed for overloading these special methods in subclasses. Perhaps it would be better to implement the logic of addition in other method (e.g. _add()) or global function which accepts the context argument, and call it in __add__(), __sub__(), and other methods. It still not guarantee that Decimal subclasses with overloaded special methods work the same in both implementations, but it is closer to the C implementation.

cc @mdickinson

antonok-edm commented 6 months ago

It is other way: the Python implementation accepts and passes the context argument to some special methods, while the C implementation has standard signatures.

Ah, right, I managed to get it backwards :sweat_smile: Updated the original post.

rhettinger commented 6 months ago

FWIW the extra context parameter was considered an implementation detail from the outset. It was not a documented part of the public API. In practice this was worked out well and proven harmless. The users likely never see it because arithmetic dunder methods aren't typically called directly (the + operator is more convenient and it supports the fallback to __radd__).

I recommend leaving this as-is. There isn't anything worth churning all this long stable code. The referenced issue in the rp2 project is rather exotic and not typically for a user of the decimal module.

If anything, I would suggest adding a note to the docs acknowledging the implementation detail.

Perhaps it would be better to implement the logic of addition in other method (e.g. _add()) or global function which accepts the context argument, and call it in add(), sub(), and other methods

This would be a lot of code churn for almost no benefit. It would make the module even harder to read and maintain. Also, in the pure python verion, and extra layer of function call forwarding would just may the module slower.