There are a number of reasons why it might make sense to separate the absolute and relative components of relativedelta.relativedelta:

1. People are often confused by the fact that the arguments have such similar names - month vs. months - easy to miss in code review, etc.
2. The absolute arguments make it way less meaningful to compare these things, and the differences around which components change if you add/subtract things are not intuitive.
3. There is a meaningful abstraction for the absolute components - a "partial date" - e.g. "month=2, day=1".

I suggest an implementation where all the absolute components are refactored out of relativedelta and into a partialdatetime object - which itself would be composed of a partialdate and partialtime object. These would have the following properties:

• They can be fully empty, and empty partial* objects are falsy.
• They can be used to update existing dates - the interface for this is unclear, but we'll want a simple method for using them to update `datetime objects, some options:
  • Implement them as a subclass of something that can be expanded with double-stars (datetime.replace(**p_dt))
  • A method that returns a dict (datetime.replace(**p_dt.asdict()))
  • A method on partialdate that updates the datetime (partialdate.update(dt)) (not great)
  • Override __add__ to mimic the existing relativedelta interface with datetime (also not great)

The following are still open questions:

• Mutability - should they be immutable like datetime, or mutable? I'm leaning towards immutable.
• Binary operations - how should they be defined? What is partialdate(year=2017, month=2) < partialdate(year=2017, day=14)? How about partialdate(year=2017, month=2) - partialdate(year=2016)?

I'd say that they should be added as a keyword argument and main implementation for relativedelta relatively soon, and then the absolute arguments should be deprecated in the following release. Ideally we'd factor out the non-absolute components into a separate object at the same time, but I think the correct name for that object is relativedelta, so I'm not sure what we would want to call the super-object in that case.