load_only and dump_only validation

[WesDowney]

I was wondering if you would want some additional validation on this. I can put up a quick pull request if you like.

In schema.py, dump_only and load_only could be validated in the same way only and exclude are.

I ran into a scenario where I had something like MySchema(partial=True,dump_only=('name')) (note the missing comma at the end of the tuple. This passes in the string name and in the Marshmallow code, gets turned into a set. set(['a', 'e', 'm', 'n']) This happens silently. I could add the same validators that only and exclude use.

Thoughts?

[lafrech]

No objection.

Could this be considered a breaking change? (Would a user actually pass "abcde" to exclude fields "a", "b", "c", "d", and "e"?)

After a quick look, it appears to me again we need to harmonize how parameters are passed (#1244). only and exclude are treated differently. Some checks run if passed in Meta but not at init.

[WesDowney]

It doesn't seem likely to me that someone would pass a string expecting it to be chopped up into characters. The most likely thing to break for people would be if they had it silently failing, now it wouldn't be silent anymore alerting them to a bug in their code.

The docs look like they are expected to be similar. (Union[Sequence[str], Set[str]]):

class marshmallow.Schema(*, only: Union[Sequence[str], Set[str]] = None, exclude: Union[Sequence[str], Set[str]] = (), many: bool = False, context: Dict = None, load_only: Union[Sequence[str], Set[str]] = (), dump_only: Union[Sequence[str], Set[str]] = (), partial: Union[bool, Sequence[str], Set[str]] = False, unknown: str = None)[source]