What

FeatureBuilder.get_column_info() sets feature_cols=[] to signal "read all source columns" for transformation modes that compute output features from raw input. This signal was only applied for "ray" and "pandas" modes — "python" was missing.
TransformationMode.PYTHON is the default mode for @batch_feature_view. This means every default @batch_feature_view using SparkComputeEngine was broken out of the box.

Why it breaks

With feature_cols populated from the view's output schema (e.g. ["user_avg_rating", "user_review_count"]), the Spark offline store read step generates:

SELECT user_id, user_avg_rating, user_review_count, event_timestamp
FROM s3a://bucket/processed/reviews/
WHERE event_timestamp BETWEEN ...

These output feature names don't exist in the raw source data — the UDF is supposed to compute them. Spark's analyzer throws:

AnalysisException: [UNRESOLVED_COLUMN.WITH_SUGGESTION] A column or function
parameter with name 'user_avg_rating' cannot be resolved.

Fix

Added "python" to the existing mode guard in get_column_info()

[x] Verified end-to-end on OpenShift AI with Feast 0.62.0 + PySpark 3.5.3 + MinIO + Redis — 2M+ feature vectors materialized successfully after fix