SNOW-948486 Support specifying input column names for vectorized UDTF

[sfc-gh-stan]

SNOW-948486

The PR introduces a new optional parameter input_names to UDTFRegistration.register/register_file and functions.py/pandas_udtf, where we now support specifying input column names for UDTF's. Note this is essentially a NOOP for regular UDTFs since the columns are mapped to function parameters, the feature is only useful for vectorized UDTF where we rely on names to access the columns in the pandas DataFrame. The parameter is set as optional for backwards compatibility, if unspecified, the default column names will be ARG1, ARG2, etc.

As a result, we now infer the column names from schema in RelationalGroupedDataFrame.applyInPandas.

[sfc-gh-sfan]

Could you elaborate on how this API change can work together with the behavior change? i.e. What happens if we release before the BCR server change, and what happens if we release after the BCR server change.

[sfc-gh-stan]

Could you elaborate on how this API change can work together with the behavior change? i.e. What happens if we release before the BCR server change, and what happens if we release after the BCR server change.

The server change has already been released and enabled in all deployments by default. The change basically allows users to specify the input column names from function definition command (using the signature). Previously the function signature will be ignored and users had to set columns like pdf.columns = ["col1", "col2"] in the function body.
Note that specifying column names is essentially a NOOP for regular UDTF's, since we just use function parameters to access the input columns, so client never supported specifying column names for UDTF input. That's why this change is required to accommodate the BCR.
Please see the changed doc tests for how BCR impacts vectorized UDTF registration.

[sfc-gh-mkeller]

🚢 🚢