Hi all,

Dataclasses provide a unique way of creating parameter bags with decent typing and validation options. However, they currently lack a meaningful way of detecting a missing value as a dataclass interface consumer.

Current behavior allows us to specify defaults per property, but defining a default value has consequences for the reusability of the dataclass, especially when it comes to inheritance patterns. Aside from this effect, littering the code with `str_field: str = MISSING_STR` does not really provide a clean and readable interface.

Therefore, I would like to propose an extra option on both the dataclasses.dataclass decorator as well as a dataclasses.field boolean argument called `allow_missing`.

Example:

@dataclass(allow_missing=True)

class UserBag:

  user_id: UUID

bag = UserBag() # no exception with regards to missing args

bag.user_id # type: Missing[UUID] ?

if not bag.user_id: # sentinel value evaluates to false

  ...

if dataclasses.ismissing(bag.user_id): # or more explicit

  ...

Alternatively, as api on a per-field-basis:

@dataclass

class UserBag:

  user_id: UUID = field(allow_missing=True)

As a bonus, one would imagine the `asdict` method has an argument to convert a dataclass instance to a dictionary without the missing properties (example: `asdict(d, exclude_missing=True)`).

Generation of code could stay largely the same as it currently is, with only minor differences:

- type hinting of constructor args should provide a sentinel default value that adheres to the typing of the original field, something along the lines of `__init__(self, arg = Missing[str])` 

- exception currently raised if a Missing[T] value is detected after construction should be silenced if the field is marked as missable.

- current sentinel MISSING value evaluates to True when cast as boolean

I would love to have your feedback on this.

Best,