Skip to content

perf: Online feature response construction in a single pass over read rows #6186

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ntkathole wants to merge 1 commit into
base: master
Choose a base branch
from
Open

perf: Online feature response construction in a single pass over read rows #6186

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 2 additions & 7 deletions sdk/python/feast/feature_store.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3064,18 +3064,13 @@ def _retrieve_from_online_store_v2(
entity_key_dict,
)

feature_data = utils._convert_rows_to_protobuf(
requested_features=features_to_request,
read_rows=list(zip(datevals, list_of_feature_dicts)),
)

online_features_response = GetOnlineFeaturesResponse(results=[])
utils._populate_response_from_feature_data(
feature_data=feature_data,
requested_features=features_to_request,
read_rows=list(zip(datevals, list_of_feature_dicts)),
indexes=idxs,
online_features_response=online_features_response,
full_feature_names=False,
requested_features=features_to_request,
table=table,
output_len=output_len,
include_feature_view_version_metadata=include_feature_view_version_metadata,
Expand Down
18 changes: 4 additions & 14 deletions sdk/python/feast/infra/online_stores/online_store.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,17 +220,12 @@ def get_online_features(
requested_features=requested_features,
)

feature_data = utils._convert_rows_to_protobuf(
requested_features, read_rows
)

# Populate the result_rows with the Features from the OnlineStore inplace.
utils._populate_response_from_feature_data(
feature_data,
requested_features,
read_rows,
idxs,
online_features_response,
full_feature_names,
requested_features,
table,
output_len,
include_feature_view_version_metadata,
Expand Down Expand Up @@ -356,17 +351,12 @@ async def query_table(table, requested_features):
for (idxs, read_rows, output_len), (table, requested_features) in zip(
all_responses, grouped_refs
):
feature_data = utils._convert_rows_to_protobuf(
requested_features, read_rows
)

# Populate the result_rows with the Features from the OnlineStore inplace.
utils._populate_response_from_feature_data(
feature_data,
requested_features,
read_rows,
idxs,
online_features_response,
full_feature_names,
requested_features,
table,
output_len,
include_feature_view_version_metadata,
Expand Down
242 changes: 86 additions & 156 deletions sdk/python/feast/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -666,30 +666,6 @@ def _group_feature_refs(
return fvs_result, odfvs_result


def construct_response_feature_vector(
values_vector: Iterable[Any],
statuses_vector: Iterable[Any],
timestamp_vector: Iterable[Any],
mapping_indexes: Iterable[List[int]],
output_len: int,
) -> GetOnlineFeaturesResponse.FeatureVector:
values_output: Iterable[Any] = [None] * output_len
statuses_output: Iterable[Any] = [None] * output_len
timestamp_output: Iterable[Any] = [None] * output_len

for i, destinations in enumerate(mapping_indexes):
for idx in destinations:
values_output[idx] = values_vector[i] # type: ignore[index]
statuses_output[idx] = statuses_vector[i] # type: ignore[index]
timestamp_output[idx] = timestamp_vector[i] # type: ignore[index]

return GetOnlineFeaturesResponse.FeatureVector(
values=values_output,
statuses=statuses_output,
event_timestamps=timestamp_output,
)


def _apply_aggregations_to_response(
response_data: Union[pyarrow.Table, Dict[str, List[Any]]],
aggregations,
Expand Down Expand Up @@ -1130,115 +1106,6 @@ def ensure_request_data_values_exist(
raise RequestDataNotFoundInEntityRowsException(feature_names=missing_features)


def _populate_response_from_feature_data(
feature_data: Iterable[
Tuple[
Iterable[Timestamp], Iterable["FieldStatus.ValueType"], Iterable[ValueProto]
]
],
indexes: Iterable[List[int]],
online_features_response: GetOnlineFeaturesResponse,
full_feature_names: bool,
requested_features: Iterable[str],
table: "FeatureView",
output_len: int,
include_feature_view_version_metadata: bool = False,
):
"""Populate the GetOnlineFeaturesResponse with feature data.

This method assumes that `_read_from_online_store` returns data for each
combination of Entities in `entity_rows` in the same order as they
are provided.

Args:
feature_data: A list of data in Protobuf form which was retrieved from the OnlineStore.
indexes: A list of indexes which should be the same length as `feature_data`. Each list
of indexes corresponds to a set of result rows in `online_features_response`.
online_features_response: The object to populate.
full_feature_names: A boolean that provides the option to add the feature view prefixes to the feature names,
changing them from the format "feature" to "feature_view__feature" (e.g., "daily_transactions" changes to
"customer_fv__daily_transactions").
requested_features: The names of the features in `feature_data`. This should be ordered in the same way as the
data in `feature_data`.
table: The FeatureView that `feature_data` was retrieved from.
output_len: The number of result rows in `online_features_response`.
"""
# Add the feature names to the response.
# Use name_to_use() which includes version tag (e.g. "fv@v2") when a
# version-qualified ref was used, so multi-version queries produce
# distinct column names like "fv@v1__feat" and "fv@v2__feat".
table_name = table.projection.name_to_use()
clean_table_name = table.projection.name_alias or table.projection.name
requested_feature_refs = [
f"{table_name}__{feature_name}" if full_feature_names else feature_name
for feature_name in requested_features
]
online_features_response.metadata.feature_names.val.extend(requested_feature_refs)

# Add version metadata if requested
if include_feature_view_version_metadata:
# Check if this feature view already exists in metadata to avoid duplicates
existing_names = [
fvm.name for fvm in online_features_response.metadata.feature_view_metadata
]
if clean_table_name not in existing_names:
fv_metadata = online_features_response.metadata.feature_view_metadata.add()
fv_metadata.name = clean_table_name
# Extract version from the table's current_version_number attribute
fv_metadata.version = getattr(table, "current_version_number", 0) or 0

# Process each feature vector in a single pass
for timestamp_vector, statuses_vector, values_vector in feature_data:
response_vector = construct_response_feature_vector(
values_vector, statuses_vector, timestamp_vector, indexes, output_len
)
online_features_response.results.append(response_vector)


def _populate_response_from_feature_data_v2(
feature_data: Iterable[
Tuple[
Iterable[Timestamp], Iterable["FieldStatus.ValueType"], Iterable[ValueProto]
]
],
indexes: Iterable[List[int]],
online_features_response: GetOnlineFeaturesResponse,
requested_features: Iterable[str],
output_len: int,
):
"""Populate the GetOnlineFeaturesResponse with feature data.

This method assumes that `_read_from_online_store` returns data for each
combination of Entities in `entity_rows` in the same order as they
are provided.

Args:
feature_data: A list of data in Protobuf form which was retrieved from the OnlineStore.
indexes: A list of indexes which should be the same length as `feature_data`. Each list
of indexes corresponds to a set of result rows in `online_features_response`.
online_features_response: The object to populate.
full_feature_names: A boolean that provides the option to add the feature view prefixes to the feature names,
changing them from the format "feature" to "feature_view__feature" (e.g., "daily_transactions" changes to
"customer_fv__daily_transactions").
requested_features: The names of the features in `feature_data`. This should be ordered in the same way as the
data in `feature_data`.
output_len: The number of result rows in `online_features_response`.
"""
# Add the feature names to the response.
requested_feature_refs = [(feature_name) for feature_name in requested_features]
online_features_response.metadata.feature_names.val.extend(requested_feature_refs)

timestamps, statuses, values = zip(*feature_data)

# Populate the result with data fetched from the OnlineStore
# which is guaranteed to be aligned with `requested_features`.
for timestamp_vector, statuses_vector, values_vector in feature_data:
response_vector = construct_response_feature_vector(
values_vector, statuses_vector, timestamp_vector, indexes, output_len
)
online_features_response.results.append(response_vector)


def _convert_entity_key_to_proto_to_dict(
entity_key_vals: List[EntityKeyProto],
) -> Dict[str, List[ValueProto]]:
Expand Down Expand Up @@ -1612,36 +1479,99 @@ def _get_entity_key_protos(
return entity_key_protos


def _convert_rows_to_protobuf(
def _populate_response_from_feature_data(
requested_features: List[str],
read_rows: List[Tuple[Optional[datetime], Optional[Dict[str, ValueProto]]]],
) -> List[Tuple[List[Timestamp], List["FieldStatus.ValueType"], List[ValueProto]]]:
n_rows = len(read_rows)
indexes: Iterable[List[int]],
online_features_response: GetOnlineFeaturesResponse,
full_feature_names: bool,
table: "FeatureView",
output_len: int,
include_feature_view_version_metadata: bool = False,
):
"""Populate the GetOnlineFeaturesResponse from raw online_read rows.

Converts raw rows from the OnlineStore into protobuf FeatureVectors and
appends them to the response. This method assumes that ``online_read``
returns data for each unique entity in the same order as ``indexes``.

Args:
requested_features: The names of the features to extract from
each row. Determines the order of FeatureVectors in the response.
read_rows: Raw output from ``OnlineStore.online_read`` — a list of
``(event_timestamp, feature_dict)`` tuples, one per unique entity.
``feature_dict`` may be ``None`` when the entity is not found.
indexes: A tuple of lists that maps each unique entity (by position
in ``read_rows``) to one or more output positions in the response.
Used to fan-out deduplicated reads back to the original request rows.
online_features_response: The protobuf response object to populate.
full_feature_names: If True, feature names are prefixed with the
feature view name (e.g. ``"driver_fv__trips_today"``).
table: The FeatureView that ``read_rows`` was retrieved from.
output_len: Total number of result rows in the response.
include_feature_view_version_metadata: If True, version metadata
for the feature view is added to the response.
"""
n_features = len(requested_features)

table_name = table.projection.name_to_use()
clean_table_name = table.projection.name_alias or table.projection.name
feature_refs = [
f"{table_name}__{fn}" if full_feature_names else fn for fn in requested_features
]
online_features_response.metadata.feature_names.val.extend(feature_refs)

if include_feature_view_version_metadata:
existing_names = [
fvm.name for fvm in online_features_response.metadata.feature_view_metadata
]
if clean_table_name not in existing_names:
fv_metadata = online_features_response.metadata.feature_view_metadata.add()
fv_metadata.name = clean_table_name
fv_metadata.version = getattr(table, "current_version_number", 0) or 0

null_value = ValueProto()
null_status = FieldStatus.NOT_FOUND
present_status = FieldStatus.PRESENT
null_ts = Timestamp()
PRESENT = FieldStatus.PRESENT
NOT_FOUND = FieldStatus.NOT_FOUND

# Pre-compute timestamps once per entity (not per feature)
# This reduces O(features * entities) to O(entities) for timestamp conversion
row_timestamps = []
row_ts_protos = []
for row_ts, _ in read_rows:
ts_proto = Timestamp()
ts = Timestamp()
if row_ts is not None:
ts_proto.FromDatetime(row_ts)
row_timestamps.append(ts_proto)

requested_features_vectors = []
for feature_name in requested_features:
ts_vector = list(row_timestamps) # Shallow copy of pre-computed timestamps
status_vector = [null_status] * n_rows
value_vector = [null_value] * n_rows
for idx, (_, feature_data) in enumerate(read_rows):
if (feature_data is not None) and (feature_name in feature_data):
status_vector[idx] = present_status
value_vector[idx] = feature_data[feature_name]
requested_features_vectors.append((ts_vector, status_vector, value_vector))
return requested_features_vectors
ts.FromDatetime(row_ts)
row_ts_protos.append(ts)

ts_template = [null_ts] * output_len
indexes_tuple = tuple(indexes)
for row_idx, destinations in enumerate(indexes_tuple):
ts = row_ts_protos[row_idx]
for out_idx in destinations:
ts_template[out_idx] = ts

feat_values = [[null_value] * output_len for _ in range(n_features)]
feat_statuses = [[NOT_FOUND] * output_len for _ in range(n_features)]

feat_idx_map = {name: i for i, name in enumerate(requested_features)}
for row_idx, destinations in enumerate(indexes_tuple):
_, feature_data = read_rows[row_idx]
if feature_data is None:
continue
for feat_name, feat_val in feature_data.items():
f_idx = feat_idx_map.get(feat_name)
if f_idx is not None:
for out_idx in destinations:
feat_values[f_idx][out_idx] = feat_val
feat_statuses[f_idx][out_idx] = PRESENT

for f_idx in range(n_features):
online_features_response.results.append(
GetOnlineFeaturesResponse.FeatureVector(
values=feat_values[f_idx],
statuses=feat_statuses[f_idx],
event_timestamps=list(ts_template),
)
)


def has_all_tags(
Expand Down
Loading
Loading