linkml-multivalued-default

An experiment in generating an empty list default for multivalued slots in a LinkML model.

Background

It is difficult or impossible in LinkML to specify that you want an empty list as the default value for a slot that is both required and multivalued. This repository contains some materials to help demonstrate this, as well as a workaround.

For more context, see the following issues:

• dandi/dandi-archive#2150
• dandi/dandi-schema#286

Setup

1. If you are using Nix flakes, run nix develop to install dependencies. Otherwise, ensure you have Python 3.10 installed.

2. Create a virtual environment with python -m venv venv and activate it with . ./venv/bin/activate.

3. Install the Python dependencies with pip install -r requirements.txt.

Demonstration of non-optional list types with default value.

1. Generate a Pydantic model by running gen-pydantic personinfo_busted.yaml >personinfo_busted.py. Note that setting both multivalued=true and required=true correctly infers the type of the aliases3 field as List[str] (rather than Optional[List[str]] as for aliases2, which is not required), but that it is not possible to set a default value of [] directly (the Pydantic generator does not have a way to encode list values in the ifabsent attribute; attempts to do so generate the string "[]").

2. Generate another Pydantic model by running gen-pydantic personinfo_workaround.yaml >personinfo_workaround.py. Note that this model sets a globally unique and incorrectly typed default on aliases3.

3. Run gen-pydantic personinfo_workaround.yaml | sed s/\"aliases3dummy\"/[]/ >personinfo_workaround.py. Note that now the definition of aliases3 correctly has an empty list default value.

4. Fire up Python and run this script:

   from personinfo_workaround import Person
   p = Person(name2='Ada Byron', aliases2=['Lady Byron'])
   p

   Note that p contains the expected values for all fields, including a default [] for aliases3.

The need to post-process like this is unfortunate, but seems to be necessary due to unresolved issues of defaults for multivalued slots in LinkML ([1], [2]).

Conclusion

It is possible to generate correct Pydantic models from LinkML schemata featuring required, multivalued attributes with a list-valued default, but it currently requires a transformation of the generated code.