Summary

Fixes #6099 - Ensures offline retrieval honors ODFV feature projection, matching online retrieval behavior.

Problem

When requesting a subset of features from an OnDemandFeatureView:

• Online retrieval ✅ Returns only requested features
• Offline retrieval ❌ Returns ALL ODFV output features (before this fix)

This caused schema mismatches between training and serving pipelines.

Solution

Modified RetrievalJob.to_arrow() in offline_store.py to:

1. Parse requested features from metadata.features
2. Build a mapping of ODFV name → requested feature names
3. Filter ODFV transformation output to only include requested columns

Example

Before this fix:

features = ["my_odfv:feature_a"]
offline_result = store.get_historical_features(features=features, ...)
# Columns: driver_id, event_timestamp, feature_a, feature_b, feature_c ❌

After this fix:

features = ["my_odfv:feature_a"]
offline_result = store.get_historical_features(features=features, ...)
# Columns: driver_id, event_timestamp, feature_a ✅

Changes

Modified: sdk/python/feast/infra/offline_stores/offline_store.py

• Updated RetrievalJob.to_arrow() method (lines 140-184)
• Added filtering logic for ODFV output projection
• Maintains backward compatibility

Added: Test in sdk/python/tests/integration/offline_store/test_universal_historical_retrieval.py

• test_odfv_projection() - Comprehensive test verifying:
  • Single feature request returns only that feature
  • Multiple feature request returns only requested features
  • Unrequested features are NOT included
  • Offline and online retrieval have consistent behavior
• Parametrized for both full_feature_names=True and False

Testing

The new test test_odfv_projection verifies:

1. ✅ Requesting 1 out of 3 ODFV features → returns only that 1 feature
2. ✅ Requesting 2 out of 3 ODFV features → returns only those 2 features
3. ✅ Unrequested features are NOT included in the result
4. ✅ Offline and online retrieval return consistent schemas

Backward Compatibility

• ✅ Falls back to old behavior if metadata is unavailable
• ✅ No breaking changes to existing functionality
• ✅ Only affects ODFV feature projection

Impact

This fix ensures:

• ✅ Consistent behavior between online and offline retrieval
• ✅ No schema mismatches in ML pipelines
• ✅ More efficient - doesn't compute/return unnecessary features
• ✅ Matches user expectations - returns exactly what was requested