ModelSerializer inconsistently serializes UUID foreign keys

[r-thomson]

I'm working in a REST Framework app that's using UUIDs for primary keys, and I've run into some unexpected behavior with the way ModelSerializer handles foreign keys. I think it's worth considering changing this behavior to make things more consistent.

Example

Model File

from django.db import models

class PrimaryModel(models.Model):
    id = models.UUIDField(primary_key=True)
    related = models.ForeignKey('RelatedModel', on_delete=models.DO_NOTHING)

class RelatedModel(models.Model):
    id = models.UUIDField(primary_key=True)

Serializer File

from rest_framework import serializers

class ExampleSerializer(serializers.ModelSerializer):
    class Meta:
        model = PrimaryModel
        fields = ('id', 'related', 'related_id')

If we look at the fields that get generated by ModelSerializer, we can see that the fields are mapped as so:

• id (models.UUIDField) becomes serializers.UUIDField
• related becomes serializers.PrimaryKeyRelatedField
• related_id becomes serializers.ReadOnlyField

ExampleSerializer():
    id = UUIDField(validators=[...])
    related = PrimaryKeyRelatedField(queryset=RelatedModel.objects.all())
    related_model = ReadOnlyField()

If we serialize a PrimaryModel instance, we can see that we actually get different types in the response: id becomes a str, while the rest become UUIDs.

{
    'id': 'a04f3708-35b7-4ae6-98c5-56e7a598de85',
    'related': UUID('1e5aa453-b351-4d56-a2c1-cf5697a4892b'),
    'related_id': UUID('6100547b-f0ac-4ab2-be8c-23076fee2741')
}

In practice, JSONRenderer will convert any UUID object to strings, so it makes no difference for the public API. However, it definitely introduces some friction when working with the serializer data directly, like when writing tests.

Suggested Solution

I think the most sensible behavior here would be for the pk_field kwarg on PrimaryKeyRelatedField to be automatically set to serializers.UUIDField if the foreign key's type is a UUID, unless the user explicitly sets it to something else. This could be controlled by a setting to avoid breaking changes.

While the behavior of related_id here is still a bit of a gotcha, I think it ultimately makes sense given the behavior of ReadOnlyField.

[auvipy]

That looks interesting! can you start a PR with a failing test case?

[r-thomson]

See #8791

I can also take a crack at implementation, if you’d like

[auvipy]

you love to see that as well. considering that it will fix the bug and not introduce any new API surface and won't create major regression

[r-thomson]

#8791 updated 👍