@average-vr-56795: I notice I keep adding

__new__

overrides to datatypes to allow setting defaults arg values… Do you have any thoughts for how to pull that nicely into the superclass?

i think my first thought is that my

__new__

overrides will typically do some sort of validation as well, after doing the

super()

call -- and that this can be distinguished from default args values, which typically are determined before doing the

super()

call. so one way this can be improved is just to make a separate

classmethod

which does the sometimes-complex

super(MyType, cls).__new__(cls, *args, **kwargs)

call for you, and then making

__new__

methods is less overhead. i think another probably-better alternative would be to have one

classmethod

you can override and return a dict to set default values, and another

classmethod

you can override to perform validation of the object post-construction (where the return value is ignored (or maybe not)). i like the idea of having a declarative way to denote at least default values for arguments, and then have the `classmethod`s described above being considered a more low-level way to do so. the most obvious way i can think of is to just allow another optional element of the field tuple which contains the default value for the field. this also means the body of the

datatype

method knows about the default values, and could maybe do something with them if it wants to. i don't see a huge risk of mixing up arguments there, but one way to make that more ergonomic would be to also define a

namedtuple

class with a 1-2 letter class name to pass into the

datatype()

call just so people can use named args if they feel like it (code using these named args might look less intimidating). i agree that the proposed two `classmethod`s above (one for default args, one for validation post-construction) could be considered unnecessary when you can just override

__new__

and do both at once, but i also like the idea of explicitly splitting out default argument selection and object validation, and i think this will lead to more terse, readable code (spoilers: i did this in R a while ago and it worked but nobody was there to critique me).