Modelkit 0.1 migration note

Modelkit relies on pydantic as part of its validation process.

Modelkit 0.1 and onwards will be shipped with pydantic 2, which comes with significant performance improvements at the cost of breaking changes.

Details on how to migrate to pydantic 2 are available in the corresponding migration guide: https://docs.pydantic.dev/latest/migration/

Installation

To install the brand new stable release of modelkit:

pip install modelkit --upgrade

Known breaking changes

Some breaking changes are arising while upgrading to pydantic 2 and the new modelkit 0.1. Here is a brief, rather exhaustive, list of the encountered issues or dropped features.

Drop: implicit pydantic model conversion

With pydantic < 2 and modelkit < 0.1, the following pattern was authorized (even though not advised) due to implicit conversions between pydantic models:

import modelkit
import pydantic
import typing

class OutputItem(pydantic.BaseModel):
    x: int

class AnotherOutputItem(pydantic.BaseModel):
    x: int

class MyModel(modelkit.Model[int, OutputItem]):
    def _predict(self, item):
        return AnotherOutputItem(x=item)

model = MyModel()
model(1)  # raises!

This pattern is no longer allowed.

However, here are the fixes: - directly build the right output pydantic Model (here: OutputItem) - directly use dicts to benefit from the dict to model conversion from pydantic and modelkit (or via .model_dump())

Drop: model validation deactivation

The MODELKIT_ENABLE_VALIDATION environment variable (or the enable_validation parameter of the LibrarySettings) which allowed one to deactivate validation if set to False was removed.

This feature has worked for pydantic < 2 for rather simple pydantic models but not complex ones with nested structures (see: https://github.com/Cornerstone-OnDemand/modelkit/pull/8). However, it still is an open question in pydantic 2, whether to allow recursive construction of models without validation (see: https://github.com/pydantic/pydantic/issues/8084). Due to the fact pydantic 2 brings heavy performance improvements, this feature has not been re-implemented.

Fixes: None, just prepare to have your inputs / outputs validated :)

Development Workflows

modelkit 0.1 (and forward) changes will be pushed to the main branch.

For projects that have not migrated, modelkit 0.0 will continue to receive maintenance on the v0.0-maintenance branch. Releases on PyPI and manual tags will adhere to the usual process.

To prevent your project from automatically upgrading to the new modelkit 0.1 upon its stable release, you can enforce an upper bound constraint in your requirements, e.g.: modelkit<0.1