Optimize _convert_rows_to_protobuf: eliminate redundant timestamp conversions

[abhijeet-dhumal]

The _convert_rows_to_protobuf function in sdk/python/feast/utils.py performs redundant timestamp conversions. The FromDatetime() call is executed once per feature per entity, when it should only be executed once per entity.. For 50 features × 500 entities, this means 25,000 calls instead of 500.

Benchmark Results

Profile	Features	Entities	p99 Latency	protobuf_convert Time
Small	3	10	0.4ms	0.04ms (10%)
Medium	5	100	14.9ms	0.5ms (10%)
Large	20	500	49.9ms	10.9ms (28%)
Stress	50	500	115.4ms	26.8ms (27%)

At scale, protobuf_convert consumes ~27% of request time.

Root Cause

sdk/python/feast/utils.py lines 1426-1436:

for feature_name in requested_features:      # 50 iterations
    for idx, read_row in enumerate(read_rows):  # 500 iterations
        row_ts_proto.FromDatetime(row_ts)       # 25,000 calls (should be 500)

Specifications

• Version: latest v0.60.0 branch
• Platform: Local

Possible Solution

# Pre-compute once per entity
row_timestamps = [Timestamp() for _ in read_rows]
for idx, (row_ts, _) in enumerate(read_rows):
    if row_ts is not None:
        row_timestamps[idx].FromDatetime(row_ts)

# Reference pre-computed timestamps
for feature_name in requested_features:
    ts_vector = row_timestamps[:]  # Shallow copy

Expected Impact

• Complexity: O(features × entities) → O(entities) for timestamp conversion
• Estimated improvement: 15-25% latency reduction for large requests
• Affected stores: All (DynamoDB, Redis, PostgreSQL, SQLite)