Skip to content

Commit 113fb04

Browse files
ntkatholentkathole
committed
perf: Online feature response construction in a single pass over read rows
Signed-off-by: ntkathole <nikhilkathole2683@gmail.com>
1 parent 080a9b5 commit 113fb04

5 files changed

Lines changed: 329 additions & 316 deletions

File tree

‎sdk/python/feast/feature_store.py‎

Lines changed: 2 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -3064,18 +3064,13 @@ def _retrieve_from_online_store_v2(
30643064
entity_key_dict,
30653065
)
30663066

3067-
feature_data = utils._convert_rows_to_protobuf(
3068-
requested_features=features_to_request,
3069-
read_rows=list(zip(datevals, list_of_feature_dicts)),
3070-
)
3071-
30723067
online_features_response = GetOnlineFeaturesResponse(results=[])
30733068
utils._populate_response_from_feature_data(
3074-
feature_data=feature_data,
3069+
requested_features=features_to_request,
3070+
read_rows=list(zip(datevals, list_of_feature_dicts)),
30753071
indexes=idxs,
30763072
online_features_response=online_features_response,
30773073
full_feature_names=False,
3078-
requested_features=features_to_request,
30793074
table=table,
30803075
output_len=output_len,
30813076
include_feature_view_version_metadata=include_feature_view_version_metadata,

‎sdk/python/feast/infra/online_stores/online_store.py‎

Lines changed: 4 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -220,17 +220,12 @@ def get_online_features(
220220
requested_features=requested_features,
221221
)
222222

223-
feature_data = utils._convert_rows_to_protobuf(
224-
requested_features, read_rows
225-
)
226-
227-
# Populate the result_rows with the Features from the OnlineStore inplace.
228223
utils._populate_response_from_feature_data(
229-
feature_data,
224+
requested_features,
225+
read_rows,
230226
idxs,
231227
online_features_response,
232228
full_feature_names,
233-
requested_features,
234229
table,
235230
output_len,
236231
include_feature_view_version_metadata,
@@ -356,17 +351,12 @@ async def query_table(table, requested_features):
356351
for (idxs, read_rows, output_len), (table, requested_features) in zip(
357352
all_responses, grouped_refs
358353
):
359-
feature_data = utils._convert_rows_to_protobuf(
360-
requested_features, read_rows
361-
)
362-
363-
# Populate the result_rows with the Features from the OnlineStore inplace.
364354
utils._populate_response_from_feature_data(
365-
feature_data,
355+
requested_features,
356+
read_rows,
366357
idxs,
367358
online_features_response,
368359
full_feature_names,
369-
requested_features,
370360
table,
371361
output_len,
372362
include_feature_view_version_metadata,

‎sdk/python/feast/utils.py‎

Lines changed: 86 additions & 156 deletions
Original file line numberDiff line numberDiff line change
@@ -666,30 +666,6 @@ def _group_feature_refs(
666666
return fvs_result, odfvs_result
667667

668668

669-
def construct_response_feature_vector(
670-
values_vector: Iterable[Any],
671-
statuses_vector: Iterable[Any],
672-
timestamp_vector: Iterable[Any],
673-
mapping_indexes: Iterable[List[int]],
674-
output_len: int,
675-
) -> GetOnlineFeaturesResponse.FeatureVector:
676-
values_output: Iterable[Any] = [None] * output_len
677-
statuses_output: Iterable[Any] = [None] * output_len
678-
timestamp_output: Iterable[Any] = [None] * output_len
679-
680-
for i, destinations in enumerate(mapping_indexes):
681-
for idx in destinations:
682-
values_output[idx] = values_vector[i] # type: ignore[index]
683-
statuses_output[idx] = statuses_vector[i] # type: ignore[index]
684-
timestamp_output[idx] = timestamp_vector[i] # type: ignore[index]
685-
686-
return GetOnlineFeaturesResponse.FeatureVector(
687-
values=values_output,
688-
statuses=statuses_output,
689-
event_timestamps=timestamp_output,
690-
)
691-
692-
693669
def _apply_aggregations_to_response(
694670
response_data: Union[pyarrow.Table, Dict[str, List[Any]]],
695671
aggregations,
@@ -1130,115 +1106,6 @@ def ensure_request_data_values_exist(
11301106
raise RequestDataNotFoundInEntityRowsException(feature_names=missing_features)
11311107

11321108

1133-
def _populate_response_from_feature_data(
1134-
feature_data: Iterable[
1135-
Tuple[
1136-
Iterable[Timestamp], Iterable["FieldStatus.ValueType"], Iterable[ValueProto]
1137-
]
1138-
],
1139-
indexes: Iterable[List[int]],
1140-
online_features_response: GetOnlineFeaturesResponse,
1141-
full_feature_names: bool,
1142-
requested_features: Iterable[str],
1143-
table: "FeatureView",
1144-
output_len: int,
1145-
include_feature_view_version_metadata: bool = False,
1146-
):
1147-
"""Populate the GetOnlineFeaturesResponse with feature data.
1148-
1149-
This method assumes that `_read_from_online_store` returns data for each
1150-
combination of Entities in `entity_rows` in the same order as they
1151-
are provided.
1152-
1153-
Args:
1154-
feature_data: A list of data in Protobuf form which was retrieved from the OnlineStore.
1155-
indexes: A list of indexes which should be the same length as `feature_data`. Each list
1156-
of indexes corresponds to a set of result rows in `online_features_response`.
1157-
online_features_response: The object to populate.
1158-
full_feature_names: A boolean that provides the option to add the feature view prefixes to the feature names,
1159-
changing them from the format "feature" to "feature_view__feature" (e.g., "daily_transactions" changes to
1160-
"customer_fv__daily_transactions").
1161-
requested_features: The names of the features in `feature_data`. This should be ordered in the same way as the
1162-
data in `feature_data`.
1163-
table: The FeatureView that `feature_data` was retrieved from.
1164-
output_len: The number of result rows in `online_features_response`.
1165-
"""
1166-
# Add the feature names to the response.
1167-
# Use name_to_use() which includes version tag (e.g. "fv@v2") when a
1168-
# version-qualified ref was used, so multi-version queries produce
1169-
# distinct column names like "fv@v1__feat" and "fv@v2__feat".
1170-
table_name = table.projection.name_to_use()
1171-
clean_table_name = table.projection.name_alias or table.projection.name
1172-
requested_feature_refs = [
1173-
f"{table_name}__{feature_name}" if full_feature_names else feature_name
1174-
for feature_name in requested_features
1175-
]
1176-
online_features_response.metadata.feature_names.val.extend(requested_feature_refs)
1177-
1178-
# Add version metadata if requested
1179-
if include_feature_view_version_metadata:
1180-
# Check if this feature view already exists in metadata to avoid duplicates
1181-
existing_names = [
1182-
fvm.name for fvm in online_features_response.metadata.feature_view_metadata
1183-
]
1184-
if clean_table_name not in existing_names:
1185-
fv_metadata = online_features_response.metadata.feature_view_metadata.add()
1186-
fv_metadata.name = clean_table_name
1187-
# Extract version from the table's current_version_number attribute
1188-
fv_metadata.version = getattr(table, "current_version_number", 0) or 0
1189-
1190-
# Process each feature vector in a single pass
1191-
for timestamp_vector, statuses_vector, values_vector in feature_data:
1192-
response_vector = construct_response_feature_vector(
1193-
values_vector, statuses_vector, timestamp_vector, indexes, output_len
1194-
)
1195-
online_features_response.results.append(response_vector)
1196-
1197-
1198-
def _populate_response_from_feature_data_v2(
1199-
feature_data: Iterable[
1200-
Tuple[
1201-
Iterable[Timestamp], Iterable["FieldStatus.ValueType"], Iterable[ValueProto]
1202-
]
1203-
],
1204-
indexes: Iterable[List[int]],
1205-
online_features_response: GetOnlineFeaturesResponse,
1206-
requested_features: Iterable[str],
1207-
output_len: int,
1208-
):
1209-
"""Populate the GetOnlineFeaturesResponse with feature data.
1210-
1211-
This method assumes that `_read_from_online_store` returns data for each
1212-
combination of Entities in `entity_rows` in the same order as they
1213-
are provided.
1214-
1215-
Args:
1216-
feature_data: A list of data in Protobuf form which was retrieved from the OnlineStore.
1217-
indexes: A list of indexes which should be the same length as `feature_data`. Each list
1218-
of indexes corresponds to a set of result rows in `online_features_response`.
1219-
online_features_response: The object to populate.
1220-
full_feature_names: A boolean that provides the option to add the feature view prefixes to the feature names,
1221-
changing them from the format "feature" to "feature_view__feature" (e.g., "daily_transactions" changes to
1222-
"customer_fv__daily_transactions").
1223-
requested_features: The names of the features in `feature_data`. This should be ordered in the same way as the
1224-
data in `feature_data`.
1225-
output_len: The number of result rows in `online_features_response`.
1226-
"""
1227-
# Add the feature names to the response.
1228-
requested_feature_refs = [(feature_name) for feature_name in requested_features]
1229-
online_features_response.metadata.feature_names.val.extend(requested_feature_refs)
1230-
1231-
timestamps, statuses, values = zip(*feature_data)
1232-
1233-
# Populate the result with data fetched from the OnlineStore
1234-
# which is guaranteed to be aligned with `requested_features`.
1235-
for timestamp_vector, statuses_vector, values_vector in feature_data:
1236-
response_vector = construct_response_feature_vector(
1237-
values_vector, statuses_vector, timestamp_vector, indexes, output_len
1238-
)
1239-
online_features_response.results.append(response_vector)
1240-
1241-
12421109
def _convert_entity_key_to_proto_to_dict(
12431110
entity_key_vals: List[EntityKeyProto],
12441111
) -> Dict[str, List[ValueProto]]:
@@ -1612,36 +1479,99 @@ def _get_entity_key_protos(
16121479
return entity_key_protos
16131480

16141481

1615-
def _convert_rows_to_protobuf(
1482+
def _populate_response_from_feature_data(
16161483
requested_features: List[str],
16171484
read_rows: List[Tuple[Optional[datetime], Optional[Dict[str, ValueProto]]]],
1618-
) -> List[Tuple[List[Timestamp], List["FieldStatus.ValueType"], List[ValueProto]]]:
1619-
n_rows = len(read_rows)
1485+
indexes: Iterable[List[int]],
1486+
online_features_response: GetOnlineFeaturesResponse,
1487+
full_feature_names: bool,
1488+
table: "FeatureView",
1489+
output_len: int,
1490+
include_feature_view_version_metadata: bool = False,
1491+
):
1492+
"""Populate the GetOnlineFeaturesResponse from raw online_read rows.
1493+
1494+
Converts raw rows from the OnlineStore into protobuf FeatureVectors and
1495+
appends them to the response. This method assumes that ``online_read``
1496+
returns data for each unique entity in the same order as ``indexes``.
1497+
1498+
Args:
1499+
requested_features: The names of the features to extract from
1500+
each row. Determines the order of FeatureVectors in the response.
1501+
read_rows: Raw output from ``OnlineStore.online_read`` — a list of
1502+
``(event_timestamp, feature_dict)`` tuples, one per unique entity.
1503+
``feature_dict`` may be ``None`` when the entity is not found.
1504+
indexes: A tuple of lists that maps each unique entity (by position
1505+
in ``read_rows``) to one or more output positions in the response.
1506+
Used to fan-out deduplicated reads back to the original request rows.
1507+
online_features_response: The protobuf response object to populate.
1508+
full_feature_names: If True, feature names are prefixed with the
1509+
feature view name (e.g. ``"driver_fv__trips_today"``).
1510+
table: The FeatureView that ``read_rows`` was retrieved from.
1511+
output_len: Total number of result rows in the response.
1512+
include_feature_view_version_metadata: If True, version metadata
1513+
for the feature view is added to the response.
1514+
"""
1515+
n_features = len(requested_features)
1516+
1517+
table_name = table.projection.name_to_use()
1518+
clean_table_name = table.projection.name_alias or table.projection.name
1519+
feature_refs = [
1520+
f"{table_name}__{fn}" if full_feature_names else fn for fn in requested_features
1521+
]
1522+
online_features_response.metadata.feature_names.val.extend(feature_refs)
1523+
1524+
if include_feature_view_version_metadata:
1525+
existing_names = [
1526+
fvm.name for fvm in online_features_response.metadata.feature_view_metadata
1527+
]
1528+
if clean_table_name not in existing_names:
1529+
fv_metadata = online_features_response.metadata.feature_view_metadata.add()
1530+
fv_metadata.name = clean_table_name
1531+
fv_metadata.version = getattr(table, "current_version_number", 0) or 0
16201532

16211533
null_value = ValueProto()
1622-
null_status = FieldStatus.NOT_FOUND
1623-
present_status = FieldStatus.PRESENT
1534+
null_ts = Timestamp()
1535+
PRESENT = FieldStatus.PRESENT
1536+
NOT_FOUND = FieldStatus.NOT_FOUND
16241537

1625-
# Pre-compute timestamps once per entity (not per feature)
1626-
# This reduces O(features * entities) to O(entities) for timestamp conversion
1627-
row_timestamps = []
1538+
row_ts_protos = []
16281539
for row_ts, _ in read_rows:
1629-
ts_proto = Timestamp()
1540+
ts = Timestamp()
16301541
if row_ts is not None:
1631-
ts_proto.FromDatetime(row_ts)
1632-
row_timestamps.append(ts_proto)
1633-
1634-
requested_features_vectors = []
1635-
for feature_name in requested_features:
1636-
ts_vector = list(row_timestamps) # Shallow copy of pre-computed timestamps
1637-
status_vector = [null_status] * n_rows
1638-
value_vector = [null_value] * n_rows
1639-
for idx, (_, feature_data) in enumerate(read_rows):
1640-
if (feature_data is not None) and (feature_name in feature_data):
1641-
status_vector[idx] = present_status
1642-
value_vector[idx] = feature_data[feature_name]
1643-
requested_features_vectors.append((ts_vector, status_vector, value_vector))
1644-
return requested_features_vectors
1542+
ts.FromDatetime(row_ts)
1543+
row_ts_protos.append(ts)
1544+
1545+
ts_template = [null_ts] * output_len
1546+
indexes_tuple = tuple(indexes)
1547+
for row_idx, destinations in enumerate(indexes_tuple):
1548+
ts = row_ts_protos[row_idx]
1549+
for out_idx in destinations:
1550+
ts_template[out_idx] = ts
1551+
1552+
feat_values = [[null_value] * output_len for _ in range(n_features)]
1553+
feat_statuses = [[NOT_FOUND] * output_len for _ in range(n_features)]
1554+
1555+
feat_idx_map = {name: i for i, name in enumerate(requested_features)}
1556+
for row_idx, destinations in enumerate(indexes_tuple):
1557+
_, feature_data = read_rows[row_idx]
1558+
if feature_data is None:
1559+
continue
1560+
for feat_name, feat_val in feature_data.items():
1561+
f_idx = feat_idx_map.get(feat_name)
1562+
if f_idx is not None:
1563+
for out_idx in destinations:
1564+
feat_values[f_idx][out_idx] = feat_val
1565+
feat_statuses[f_idx][out_idx] = PRESENT
1566+
1567+
for f_idx in range(n_features):
1568+
online_features_response.results.append(
1569+
GetOnlineFeaturesResponse.FeatureVector(
1570+
values=feat_values[f_idx],
1571+
statuses=feat_statuses[f_idx],
1572+
event_timestamps=list(ts_template),
1573+
)
1574+
)
16451575

16461576

16471577
def has_all_tags(

0 commit comments

Comments
 (0)