Split general component behavior from user implementation

[RobbeSneyders]

Currently users implement components by subclassing one of the provided component classes, which themselves provide a lot of general component behavior, such as data loading, validating, and writing.

Instead, we could split this by letting the user implement their components by implementing a class with a specific interface which is passed to a general component class.

Current:

from fondant import PandasTransformComponent 

class MyComponent(PandasTransformComponent):

    def setup(self, kwarg1, kwarg2):
        ...

    def transform(self, dataframe):
        ...

component = MyComponent.from_args()
component.run()

Future:

from fondant import Component

class MyComponent:

    def __init__(self, kwarg1, kwarg2):
        ...

    def transform(self, dataframe):
        ...

Component(MyComponent).run()

We can still provide an interface which the user can subclass, for instance using a Protocol.

This has some advantages:

• Component implementations can use __init__ instead of setup which will be more familiar to users
• It's easier to test the custom component implementation, as no dummy arguments need to be provided for the general component behavior (see [LLM pipeline] Language filter component #232 (comment))