RedRoserade / pydantic-v2-recursive-issue-repro

See https://github.com/pydantic/pydantic/issues/7907

Reproduction repository for RootModel parsing of unions

This repository provides a way to reproduce the issue where Union types aren't parsed properly under the context of a RootModel in Pydantic v2.

Context

The setup is rather large, but represents a realistic scenario of rich text modelling based on Slate.

The library allows nesting paragraphs and headings inside each other, so you can effectively have such data structure:

ParagraphModel(
    nodes=[
        ParagraphModel(nodes=[TextModel(...)])
    ]
)

We represent this as follows:

• A paragraph has nodes. Each node can be a TextModel, a ParagraphModel, or a HeadingModel.
• Because of the way data is modelled in Slate, we do this with two Unions, one between ParagraphModel and HeadingModel, and the other between TextModel and the previous Union (see this)

A TextModel can have ranges, which are segments of text, and marks, which hold information about suggestions, formatting, etc. There are several kinds of marks, and a fallback MarkModel that's a catch-all for everything else. All marks have discriminators except MarkModel.

Setup

• Create a virtual environment with Python 3.8 or above (tested on 3.8 and 3.11)
• Install the requirements

Running and results

When running the script:

python -m minimal.run

I expect both cases (parsing via a RootModel of the Union and via the concrete type, ParagraphModel) to produce the same exact output. However, by default, parsing via RootModel causes result_1.nodes[0].ranges[0].marks[1] to be a bare MarkModel (the fallback) rather than ExternalMarkModel (the expected type based on the discriminator).

This causes the assertion to fail:

python -m minimal.run
Via RootModel: ParagraphModel(object='block', type='paragraph', nodes=[TextModel(object='text', ranges=[TextItemModel(marks=[EmptyMarkModel(type='empty-mark'), MarkModel(type='external-mark', data=MarkDataModel(value=''))])])])
-----------------
Via ParagraphModel: ParagraphModel(object='block', type='paragraph', nodes=[TextModel(object='text', ranges=[TextItemModel(marks=[EmptyMarkModel(type='empty-mark'), ExternalMarkModel(type='external-mark', data=SuggestionExternalMarkDataModel(type=<ExternalMarkType.SUGGESTION: 'suggestion'>, kind=<TextItemSuggestionKind.REMOVE_TEXT: 'remove_text'>))])])])
-----------------
Traceback (most recent call last):
  File "/Users/val/.pyenv/versions/3.8.15/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/Users/val/.pyenv/versions/3.8.15/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/Users/val/repos/personal/pydantic-v2-slate-demo/minimal/run.py", line 50, in <module>
    main()
  File "/Users/val/repos/personal/pydantic-v2-slate-demo/minimal/run.py", line 46, in main
    assert result_1 == result_2
AssertionError

Things known to work, and other notes

The following "work" but aren't valid options:

• If ParagraphNodeModel is modified to not include itself, the model parses properly.
• If this line is removed, the model parses properly.

A patch fixes the situation by defining the recursive model after the ParagraphModel and HeadingModel types, and rebuilding these types, without modifying the actual data representation.

Modifying this Union to use smart mode (the default) causes both models to parse incorrectly (they use MarkModel).