can y'all remind be what is best practices around UDFs. I have model in mlflow that I want to use. should I: 1. load the model from mlflow in

__init__

2. load the model from mlflow in my main script and pass as an argument like below

@daft.cls
class ModelPredictor:
    """Model predictor for Daft DataFrames."""

    def __init__(self, model: lgb.Booster, train_config: TrainConfig) -> None:
        self.model = copy.deepcopy(model)
        self.train_config = train_config

Hi Garrett, you should load it in

__init__

!

@Kevin Wang one more question, what is the correct typing here (the docs have

Series

in the example)?

Argument 1 to "predict" of "ModelPredictor" has incompatible type "*list[Expression]"; expected "Series"

@daft.method.batch(return_dtype=daft.DataType.float64())
    def predict(self, *features: daft.Series):  # type: ignore
        """Predict method for Daft Series."""
        feature_df = pd.DataFrame(
            {feature.name: feature.to_arrow() for feature in features}
        )

        return daft.Series.from_numpy(
            model_predict(self.model, feature_df, self.train_config).to_numpy()
        )

This is a static type checking error and not a runtime error right?

yeah, static type checking error

Yeah not user error. Just mypy not understanding the decorator I think

makes sense, not the first time mypy + decorators didn't play nicely

Curious if you write Ray code and have had similar issues with Ray as well. They do a similar thing for

@ray.remote

on classes to create Ray Actors. Wonder if they do anything different that plays nicer with mypy

Nothing that comes to mind immediately.

Hm, that is one potential case where you may want to load it outside of

__init__

, if there is some issue with concurrent model weight loading.

I think it's just the connection pool settings we have on our mlflow server are very low