Change Summary

This is a fix for bug #9872. The annotation Sequence ignores discriminator metadata in certain cases. The generated schema has not correct python_schema where a union is presented rather than expected tagged-union. json_schema is normal.

The bug occurs in GenerateSchema._sequence_schema. In the json-or-python schema created, the underling reference of item_type_schema is the same, therefore the metadata dict is also shared. When calling apply_discriminators on the model schema, because json_schema is walked first, it has correct conversion from union to tagged-union. But the walk will pop the discriminator from metadata. Therefore when python_schema is walked, metadata is already empty, resulting in union, not tagged-union schema.

# `metadata` object has the same memory location
metadata = s.get('metadata', {})
discriminator = metadata.pop(CORE_SCHEMA_METADATA_DISCRIMINATOR_PLACEHOLDER_KEY, None)

I suspect there are more of these kinds of reference sharing problems in other parts of schema generation process. 🤔

Note the author mentioned that the bug disappears when adding model_validator to both models. This is because the procedure for applying discriminators is different. It is GenerateSchema._apply_discriminator_to_union that generates tagged-union in this case.

Related issue number

fix #9872

Checklist

• The pull request title is a good summary of the changes - it will be used in the changelog
• Unit tests for the changes exist
• Tests pass on CI
• Documentation reflects the changes where applicable
• My PR is ready to review, please add a comment including the phrase "please review" to assign reviewers