Replace type checks with type guard

[davidhopkinson26]

Tubular currently has a lot of type checking. This could be replaced with typeguard to simplify the code and reduce test burden.

[davidhopkinson26]

Closely linked with #116 as type hints required for typeguard

[davidhopkinson26]

the downside of this would be that we would no longer get the class specific error messages prefixed by self.classname(). This is quite a useful feature when running transformers in a long pipeline so may be a good justification for keeping the current error handling.

[davidhopkinson26]

question for investigation - Can we extend typeguard or something similar to check types of columns in pandas dataframes?

[adamsardar]

Now that #118 has been cracked, this ticket comes into focus.

The other big player in the argument validation world is pydantic and the @validate_call decorator. Comparing the two:

Pydantic

• ✔️ Technically a parsing library, not a validation library (see here). It makes no guarantee whatsoever about checking the form of data you give to it; instead it makes a guarantee about the form of data you get out of it.
  • A str value passed to a List[str] argument is made into a list (useful in tackling Consistent args and generic tests for columns: str or list transformers #150?)
  • A datetime in one format passed to a datetime of another format can also be handled (useful in relation to Date transformers casting dtypes if passed in as an input #142?)
• ✔️ Makes more use of Annotated field information; we can stipulate that an argument needs to be positive definite concisely and integrate with static type checkers. This is a nicer user experience.
• ✔️ We already use pydantic in our stack internally, so the effective dependency change is low
• ❌ Errors are raised as a ValidationError, not a TypeError as is currently done in the codebase. This would be a breaking change unfortunately.
• ❌ It's just a decorator - there's no smart way to add it across a module without annotating every method individually.
• ❌ validate_arguments decorator is in beta.

Typeguard

• ✔️ Has a means of validating the entirety of a module via a hook.
• ✔️ Considerably smaller library in pure python. Dependencies matter.
• ✔️ Has a narrower range of responsibilities - it checks types. I don't believe that it parses (read, casts).
• ❌ Raises a TypeCheckError rather than a TypeError as is currently done in the codebase. As above, this would be a breaking change.