Unusual behavior with is_valid

[Viicos]

With the following example:

class V1Serializer(serializers.Serializer):
    v1_required_field = serializers.CharField()

class V2Serializer(serializers.Serializer):
    v2_optional_field = serializers.CharField(required=False)

class MainSerializer(PolymorphicSerializer):
    discriminator_field = "version"
    serializer_mapping = {
        1: V1Serializer,
        2: V2Serializer,
    }

    version = serializers.ChoiceField(
        choices=[1, 2],
        default=1,
    )
    main_field = serializers.CharField(required=False)     

I get some unusual behaviors when doing:

MainSerializer(data={"version": "v1"}).is_valid()
#> KeyError: "`MainSerializer.serializer_mapping` is missing a corresponding serializer for the `<class 'rest_framework.fields.empty'>` key"

---

The is_valid override:

drf-polymorphic/drf_polymorphic/serializers.py

Lines 90 to 97 in 8166533

	def is_valid(self, *args, **kwargs):
	valid = super().is_valid(*args, **kwargs)
	extra_serializer = self._get_serializer_from_data(self.validated_data)
	if extra_serializer is None:
	return valid
	extra_valid = extra_serializer.is_valid(*args, **kwargs)
	self._errors.update(extra_serializer.errors)
	return valid and extra_valid

adds an extra layer of validation, by calling is_valid again on the mapped serializer from the discriminator. However, if the first super().is_valid() call failed, self.validated_data is set to {} by DRF, so drf-polymorphic is unable to retrieve the serializer from the mapping, which leads to the KeyError.

The question now is: why is super().is_valid() failing? In the example, the data provided to MainSerializer seems to comply, at first sight there is no reason for super().is_valid() to fail. The reason is that the DRF implementation of is_valid calls to_internal_value which is overridden in our case:

drf-polymorphic/drf_polymorphic/serializers.py

Lines 82 to 88 in 8166533

	def to_internal_value(self, data):
	default = super().to_internal_value(data)
	serializer = self._get_serializer_from_data(data)
	if serializer is None:
	return default
	extra = serializer.to_internal_value(data)
	return {**default, **extra}

and calls to_internal_value on the mapped serializer, which will run validation on each field of this serializer.

---

So it's a bit of a mess, validation is run twice but simply removing the extra logic in is_valid is not enough as we still need to run the mapped serializer validators (as to_internal_value only runs fields validators).

However, I think better exception handling could be applied so that is_valid(raise_exception=False) doesn't raise unexpected errors.