Provide option to strip feature view names from feature names #1618
Description
Is your feature request related to a problem? Please describe.
Users often have hardcode logic in model training pipelines based on specific feature names that come from their data sources. When Feast is introduced during productionization, we expect users to rename their feature names to the Feast convention feature_view__feature. This creates an unnecessary overhead for teams if they have many features.
Describe the solution you'd like
Provide a way to strip the feature_view__ prefix from feature names during retrieval and reuse feature names directly from sources.
before
training_df = store.get_historical_features(
entity_df=entity_df,
feature_refs = [
'driver_performance:conv_rate',
'driver_performance:acc_rate',
'driver_hourly_stats:avg_daily_trips'
],
)
for feature_name in training_df.columns:
print(feature_name)
driver_performance__conv_rate
driver_performance__acc_rate
driver_hourly_stats__avg_daily_trips
after
training_df = store.get_historical_features(
entity_df=entity_df,
feature_refs = [
'driver_performance:conv_rate',
'driver_performance:acc_rate',
'driver_hourly_stats:avg_daily_trips'
],
feature_names_only=True
)
for feature_name in training_df.columns:
print(feature_name)
conv_rate
acc_rate
avg_daily_trips
Any collisions in feature names would result in an exception.