Optimize non-discriminated unions based on incompatible schemas

[adriangb]

It would be nice if Pydantic automatically figured out the path of least cost to determine which item in a union we should pursue. Essentially we'd want to figure out the divergent points between two schemas and check those via a depth-first sort of validation instead of a breadth-first type validation that we currently do. Some obvious cases are:

• A literal field that diverges
• Fields that are only present in one model or some of the models
• Fields that have totally incompatible types

It does get tricky e.g. with a combination of these.

[adriangb]

#433 (comment) made me think that maybe the thing to do here is at runtime capture what made validation fail for the "wrong" choices and try that first.

In other words, given:

class Cat:
  nicknames: List[str]
  type: Literal["cat"]

class Dog:
  nicknames: List[str]
   type: Literal["dog"]

Pet = Cat | Dog

One the first run with the data {"type": "dog", "age": 2} we would do what we currently do which is try every option in order. But we would take note of the fact that we failed to validate against cat because of the type field.

So next time we run some data through we check if data["type"] validates against Cat -> type before we do other checks (e.g. nicknames).