import copy

import functools

import itertools

import logging

import os

import socket

import threading

import typing

import warnings

from collections import Counter, defaultdict

from datetime import datetime, timedelta, timezone

from pathlib import Path

from typing import (

Any,

Dict,

Iterable,

List,

Mapping,

Optional,

Sequence,

Set,

Tuple,

Union,

cast,

)

import pandas as pd

import pyarrow

from dateutil.tz import tzlocal

from google.protobuf.timestamp_pb2 import Timestamp

from feast.aggregation import aggregation_specs_to_agg_ops

from feast.constants import FEAST_FS_YAML_FILE_PATH_ENV_NAME

from feast.entity import Entity

from feast.errors import (

FeatureNameCollisionError,

FeatureViewNotFoundException,

RequestDataNotFoundInEntityRowsException,

)

from feast.field import Field

from feast.infra.compute_engines.backends.pandas_backend import PandasBackend

from feast.infra.key_encoding_utils import deserialize_entity_key

from feast.protos.feast.serving.ServingService_pb2 import (

FieldStatus,

GetOnlineFeaturesResponse,

)

from feast.protos.feast.types.EntityKey_pb2 import EntityKey as EntityKeyProto

from feast.protos.feast.types.Value_pb2 import FloatList as FloatListProto

from feast.protos.feast.types.Value_pb2 import RepeatedValue as RepeatedValueProto

from feast.protos.feast.types.Value_pb2 import Value as ValueProto

from feast.type_map import python_values_to_proto_values

from feast.types import ComplexFeastType, PrimitiveFeastType

from feast.value_type import ValueType

from feast.version import get_version

if typing.TYPE_CHECKING:

from feast.base_feature_view import BaseFeatureView

from feast.feature_service import FeatureService

from feast.feature_view import FeatureView

from feast.feature_view_projection import FeatureViewProjection

from feast.infra.registry.base_registry import BaseRegistry

from feast.on_demand_feature_view import OnDemandFeatureView

APPLICATION_NAME = "feast-dev/feast"

USER_AGENT = "{}/{}".format(APPLICATION_NAME, get_version())

@functools.lru_cache()

def _ipv6_available() -> bool:

"""Whether this process can bind AF_INET6 (cached: can't change at runtime)."""

try:

with socket.socket(socket.AF_INET6, socket.SOCK_STREAM) as sock:

sock.setsockopt(socket.IPPROTO_IPV6, socket.IPV6_V6ONLY, 0)

sock.bind(("::", 0))

return True

except OSError:

return False

def _make_dual_stack_socket(port: int, blocking: bool = False) -> socket.socket:

"""Pre-bind a dual-stack socket for a server to adopt.

A framework handed host="::" directly (uvicorn.run, loop.create_server)

sets IPV6_V6ONLY=1 on its own socket, silently dropping IPv4 clients.

Binding here with V6ONLY cleared avoids that; falls back to IPv4-only

when the host can't bind AF_INET6.

Non-blocking by default since every current caller hands the socket to

an asyncio-based server (uvicorn.Server(...).run(sockets=[sock])), which

requires that; pass blocking=True for a synchronous server instead.

"""

dual_stack = _ipv6_available()

if dual_stack:

sock = socket.socket(socket.AF_INET6, socket.SOCK_STREAM)

address: tuple = ("::", port)

else:

sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)

address = ("0.0.0.0", port)

try:

if dual_stack:

sock.setsockopt(socket.IPPROTO_IPV6, socket.IPV6_V6ONLY, 0)

sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)

sock.bind(address)

sock.listen(socket.SOMAXCONN)

sock.setblocking(blocking)

except Exception:

sock.close()

raise

return sock

def _parse_feature_or_view_ref(ref: str) -> Tuple[str, Optional[int], Optional[str]]:

"""Parse 'fv_name[@version][:feature]' into (fv_name, version_number, feature_name).

Unlike ``_parse_feature_ref``, the ':<feature>' suffix is optional, so this

also parses whole-view references (e.g. a FeatureService entry that pins a

version but selects all of the view's features). When no ':<feature>' is

present, feature_name is None. When no @version is present, version_number

is None (meaning 'latest').

Examples:

'driver_stats:trips' -> ('driver_stats', None, 'trips')

'driver_stats@v2:trips' -> ('driver_stats', 2, 'trips')

'driver_stats@latest:trips' -> ('driver_stats', None, 'trips')

'driver_stats' -> ('driver_stats', None, None)

'driver_stats@v2' -> ('driver_stats', 2, None)

"""

import re

colon_idx = ref.find(":")

if colon_idx < 0:

fv_part = ref

feature_name: Optional[str] = None

else:

fv_part = ref[:colon_idx]

feature_name = ref[colon_idx + 1 :]

at_idx = fv_part.find("@")

if at_idx < 0:

return (fv_part, None, feature_name)

fv_name = fv_part[:at_idx]

version_str = fv_part[at_idx + 1 :]

if not version_str or version_str.lower() == "latest":

return (fv_name, None, feature_name)

# Parse version number from formats like "v2", "V2"

match = re.match(r"^[vV](\d+)$", version_str)

if not match:

# Not a recognized version format — treat entire fv_part as the name

return (fv_part, None, feature_name)

return (fv_name, int(match.group(1)), feature_name)

def _parse_feature_ref(ref: str) -> Tuple[str, Optional[int], str]:

"""Parse 'fv_name@version:feature' into (fv_name, version_number, feature_name).

The ':<feature>' suffix is required; use ``_parse_feature_or_view_ref`` when

a whole-view reference (no feature) should be accepted.

If no @version is present, version_number is None (meaning 'latest').

Examples:

'driver_stats:trips' -> ('driver_stats', None, 'trips')

'driver_stats@v2:trips' -> ('driver_stats', 2, 'trips')

'driver_stats@latest:trips' -> ('driver_stats', None, 'trips')

"""

fv_name, version_num, feature_name = _parse_feature_or_view_ref(ref)

if feature_name is None:

raise ValueError(

f"Invalid feature reference '{ref}'. Expected format: '<feature_view>:<feature>' "

f"or '<feature_view>@<version>:<feature>'"

)

return (fv_name, version_num, feature_name)

def _strip_version_from_ref(ref: str) -> str:

"""Strip @version from a feature reference, returning 'fv_name:feature'.

Used to produce clean refs for output column naming.

"""

fv_name, _, feature_name = _parse_feature_ref(ref)

return f"{fv_name}:{feature_name}"

def _get_requested_on_demand_feature_views(

feature_refs: List[str],

project: str,

registry: "BaseRegistry",

allow_cache: bool = True,

) -> List["OnDemandFeatureView"]:

"""Resolve the ODFVs referenced by ``feature_refs``, honouring ``fv@vN`` pins.

Offline stores re-fetch ODFVs from the registry independently of the

already-resolved online path (``_get_feature_views_to_use``). This mirrors

that version-aware resolution so a pinned ODFV ref resolves to the pinned

snapshot; unversioned refs resolve to the promoted view (unchanged behavior).

"""

from feast.on_demand_feature_view import OnDemandFeatureView

promoted_by_name = {

odfv.name: odfv

for odfv in registry.list_on_demand_feature_views(

project, allow_cache=allow_cache

)

}

# Dedupe refs by (name, version), preserving first-seen order.

requested: Dict[Tuple[str, Optional[int]], None] = {}

for ref in feature_refs:

fv_name, version_num, _ = _parse_feature_ref(ref)

requested[(fv_name, version_num)] = None

resolved: List["OnDemandFeatureView"] = []

seen: Set[Tuple[str, Optional[int]]] = set()

for fv_name, version_num in requested:

if version_num is not None:

try:

odfv = registry.get_feature_view_by_version(

fv_name, project, version_num, allow_cache

)

except NotImplementedError:

# v0 fallback for registries without versioned lookup.

if version_num == 0:

odfv = registry.get_any_feature_view(fv_name, project, allow_cache)

else:

raise

if not isinstance(odfv, OnDemandFeatureView):

continue # ref belongs to a regular FeatureView, not an ODFV

if odfv.projection is not None:

odfv.projection.version_tag = version_num

else:

promoted = promoted_by_name.get(fv_name)

if promoted is None:

continue

odfv = promoted

key = (odfv.name, version_num)

if key in seen:

continue

seen.add(key)

resolved.append(odfv)

return resolved

def get_user_agent():

return USER_AGENT

def make_tzaware(t: datetime) -> datetime:

"""We assume tz-naive datetimes are UTC"""

if t.tzinfo is None:

return t.replace(tzinfo=timezone.utc)

else:

return t

def compute_non_entity_date_range(

feature_views: List["FeatureView"],

start_date: Optional[datetime] = None,

end_date: Optional[datetime] = None,

default_window_days: int = 30,

) -> Tuple[datetime, datetime]:

if end_date is None:

end_date = datetime.now(tz=timezone.utc)

else:

end_date = make_tzaware(end_date)

if start_date is None:

max_ttl_seconds = max(

(

int(fv.ttl.total_seconds())

for fv in feature_views

if fv.ttl and isinstance(fv.ttl, timedelta)

),

default=0,

)

start_date = end_date - timedelta(

seconds=max_ttl_seconds

if max_ttl_seconds > 0

else default_window_days * 86400

)

else:

start_date = make_tzaware(start_date)

return start_date, end_date

def make_df_tzaware(t: pd.DataFrame) -> pd.DataFrame:

"""Make all datetime type columns tzaware; leave everything else intact."""

df = t.copy() # don't modify incoming dataframe inplace

for column in df.columns:

if pd.api.types.is_datetime64_any_dtype(df[column]):

df[column] = pd.to_datetime(df[column], utc=True)

return df

def to_naive_utc(ts: datetime) -> datetime:

if ts.tzinfo is None:

return ts

else:

return ts.astimezone(timezone.utc).replace(tzinfo=None)

def maybe_local_tz(t: datetime) -> datetime:

if t.tzinfo is None:

return t.replace(tzinfo=tzlocal())

else:

return t

def get_default_yaml_file_path(repo_path: Path) -> Path:

if FEAST_FS_YAML_FILE_PATH_ENV_NAME in os.environ:

yaml_path = os.environ[FEAST_FS_YAML_FILE_PATH_ENV_NAME]

return Path(yaml_path)

else:

return repo_path / "feature_store.yaml"

def _get_requested_feature_views_to_features_dict(

feature_refs: List[str],

feature_views: List["FeatureView"],

on_demand_feature_views: List["OnDemandFeatureView"],

) -> Tuple[Dict["FeatureView", List[str]], Dict["OnDemandFeatureView", List[str]]]:

"""Create a dict of FeatureView -> List[Feature] for all requested features.

Set full_feature_names to True to have feature names prefixed by their feature view name.

"""

feature_views_to_feature_map: Dict["FeatureView", List[str]] = defaultdict(list)

on_demand_feature_views_to_feature_map: Dict["OnDemandFeatureView", List[str]] = (

defaultdict(list)

)

for ref in feature_refs:

fv_name, version_num, feature_from_ref = _parse_feature_ref(ref)

# Build the key that matches projection.name_to_use()

if version_num is not None:

feature_view_from_ref = f"{fv_name}@v{version_num}"

else:

feature_view_from_ref = fv_name

found = False

for fv in feature_views:

if fv.projection.name_to_use() == feature_view_from_ref:

found = True

feature_views_to_feature_map[fv].append(feature_from_ref)

for odfv in on_demand_feature_views:

if odfv.projection.name_to_use() == feature_view_from_ref:

found = True

on_demand_feature_views_to_feature_map[odfv].append(feature_from_ref)

if not found:

raise ValueError(f"Could not find feature view from reference {ref}")

return feature_views_to_feature_map, on_demand_feature_views_to_feature_map

def _get_column_names(

feature_view: "FeatureView", entities: List[Entity]

) -> Tuple[List[str], List[str], str, Optional[str]]:

"""

If a field mapping exists, run it in reverse on the join keys,

feature names, event timestamp column, and created timestamp column

to get the names of the relevant columns in the offline feature store table.

Returns:

Tuple containing the list of reverse-mapped join_keys,

reverse-mapped feature names, reverse-mapped event timestamp column,

and reverse-mapped created timestamp column that will be passed into

the query to the offline store.

"""

if feature_view.batch_source is None:

raise ValueError(

f"Feature view '{feature_view.name}' has no batch_source and cannot be used for offline retrieval."

)

# if we have mapped fields, use the original field names in the call to the offline store

timestamp_field = feature_view.batch_source.timestamp_field

# For feature views with aggregations, read INPUT columns from aggregations.

# This applies to StreamFeatureView, BatchFeatureView,

# or any FeatureView that has aggregations.

if hasattr(feature_view, "aggregations") and feature_view.aggregations:

# Extract unique input columns from aggregations, preserving order

feature_names = list(

dict.fromkeys(agg.column for agg in feature_view.aggregations)

)

else:

# For regular feature views, use the feature names

feature_names = [feature.name for feature in feature_view.features]

created_timestamp_column = feature_view.batch_source.created_timestamp_column

from feast.feature_view import DUMMY_ENTITY_ID

join_keys = [

entity.join_key for entity in entities if entity.join_key != DUMMY_ENTITY_ID

]

if feature_view.batch_source.field_mapping is not None:

reverse_field_mapping = {

v: k for k, v in feature_view.batch_source.field_mapping.items()

}

timestamp_field = (

reverse_field_mapping[timestamp_field]

if timestamp_field in reverse_field_mapping.keys()

else timestamp_field

)

created_timestamp_column = (

reverse_field_mapping[created_timestamp_column]

if created_timestamp_column

and created_timestamp_column in reverse_field_mapping.keys()

else created_timestamp_column

)

join_keys = [

reverse_field_mapping[col] if col in reverse_field_mapping.keys() else col

for col in join_keys

]

feature_names = [

reverse_field_mapping[col] if col in reverse_field_mapping.keys() else col

for col in feature_names

]

# We need to exclude join keys and timestamp columns from the list of features, after they are mapped to

# their final column names via the `field_mapping` field of the source.

feature_names = [

name

for name in feature_names

if name not in join_keys

and name != timestamp_field

and name != created_timestamp_column

]

return (

join_keys,

feature_names,

timestamp_field,

created_timestamp_column,

)

def _run_pyarrow_field_mapping(

table: pyarrow.Table,

field_mapping: Dict[str, str],

) -> pyarrow.Table:

# run field mapping in the forward direction

cols = table.column_names

mapped_cols = [

field_mapping[col] if col in field_mapping.keys() else col for col in cols

]

table = table.rename_columns(mapped_cols)

return table

def _get_fields_with_aliases(

fields: List[str],

field_mappings: Dict[str, str],

) -> Tuple[List[str], List[str]]:

"""

Get a list of fields with aliases based on the field mappings.

"""

for field in fields:

if "." in field and field not in field_mappings:

raise ValueError(

f"Feature {field} contains a '.' character, which is not allowed in field names. Use field mappings to rename fields."

)

fields_with_aliases = [

f"{field} AS {field_mappings[field]}" if field in field_mappings else field

for field in fields

]

aliases = [

field_mappings[field] if field in field_mappings else field for field in fields

]

return (fields_with_aliases, aliases)

def _coerce_datetime(ts):

"""

Depending on underlying time resolution, arrow to_pydict() sometimes returns pd

timestamp type (for nanosecond resolution), and sometimes you get standard python datetime

(for microsecond resolution).

While pd timestamp class is a subclass of python datetime, it doesn't always behave the

same way. We convert it to normal datetime so that consumers downstream don't have to deal

with these quirks.

"""

if isinstance(ts, pd.Timestamp):

return ts.to_pydatetime()

else:

return ts

def _columns_to_proto_values(

table: pyarrow.RecordBatch,

columns: List[Tuple[str, ValueType]],

allow_missing: bool = False,

) -> Dict[str, List[ValueProto]]:

"""Convert table columns to proto values dict.

Args:

table: PyArrow RecordBatch containing the data.

columns: List of (column_name, value_type) tuples to convert.

allow_missing: If True, skip columns not found in table. If False, raise ValueError.

Returns:

Dict mapping column names to lists of ValueProto.

"""

result: Dict[str, List[ValueProto]] = {}

for column, value_type in columns:

if column in table.column_names:

result[column] = python_values_to_proto_values(

table.column(column).to_numpy(zero_copy_only=False), value_type

)

elif not allow_missing:

raise ValueError(f"Column {column} not found in table")

return result

def _build_entity_keys(

num_rows: int,

join_keys: Dict[str, ValueType],

proto_values: Dict[str, List[ValueProto]],

) -> List[EntityKeyProto]:

"""Build entity key protos for each row.

Args:

num_rows: Number of rows to generate entity keys for.

join_keys: Dict mapping join key names to their value types.

proto_values: Dict mapping column names to lists of ValueProto values.

Returns:

List of EntityKeyProto, one per row.

"""

return [

EntityKeyProto(

join_keys=list(join_keys.keys()),

entity_values=[

proto_values[k][idx] for k in join_keys if k in proto_values

],

)

for idx in range(num_rows)

]

def _convert_arrow_to_proto(

table: Union[pyarrow.Table, pyarrow.RecordBatch],

feature_view: Union["FeatureView", "BaseFeatureView", "OnDemandFeatureView"],

join_keys: Dict[str, ValueType],

) -> List[Tuple[EntityKeyProto, Dict[str, ValueProto], datetime, Optional[datetime]]]:

# This is a workaround for isinstance(feature_view, OnDemandFeatureView), which triggers a circular import

# Check for source_request_sources or source_feature_view_projections attributes to identify ODFVs

if (

getattr(feature_view, "source_request_sources", None) is not None

or getattr(feature_view, "source_feature_view_projections", None) is not None

):

return _convert_arrow_odfv_to_proto(table, feature_view, join_keys) # type: ignore[arg-type]

else:

return _convert_arrow_fv_to_proto(table, feature_view, join_keys) # type: ignore[arg-type]

def _convert_arrow_fv_to_proto(

table: Union[pyarrow.Table, pyarrow.RecordBatch],

feature_view: "FeatureView",

join_keys: Dict[str, ValueType],

) -> List[Tuple[EntityKeyProto, Dict[str, ValueProto], datetime, Optional[datetime]]]:

# Avoid ChunkedArrays which guarantees `zero_copy_only` available.

if isinstance(table, pyarrow.Table):

table = table.to_batches()[0]

if feature_view.batch_source is None:

raise ValueError(

f"Feature view '{feature_view.name}' has no batch_source and cannot be converted to proto."

)

# TODO: This will break if the feature view has aggregations or transformations

columns = [

(field.name, field.dtype.to_value_type()) for field in feature_view.features

] + list(join_keys.items())

proto_values_by_column = _columns_to_proto_values(

table, columns, allow_missing=False

)

entity_keys = _build_entity_keys(table.num_rows, join_keys, proto_values_by_column)

# Serialize the features per row

feature_dict = {

feature.name: proto_values_by_column[feature.name]

for feature in feature_view.features

}

features = [dict(zip(feature_dict, vars)) for vars in zip(*feature_dict.values())]

# Convert event_timestamps

event_timestamps = [

_coerce_datetime(val)

for val in pd.to_datetime(

table.column(feature_view.batch_source.timestamp_field).to_numpy(

zero_copy_only=False

)

)

]

# Convert created_timestamps if they exist

if feature_view.batch_source.created_timestamp_column:

created_timestamps = [

_coerce_datetime(val)

for val in pd.to_datetime(

table.column(

feature_view.batch_source.created_timestamp_column

).to_numpy(zero_copy_only=False)

)

]

else:

created_timestamps = [None] * table.num_rows

return list(zip(entity_keys, features, event_timestamps, created_timestamps))

def _convert_arrow_odfv_to_proto(

table: Union[pyarrow.Table, pyarrow.RecordBatch],

feature_view: "OnDemandFeatureView",

join_keys: Dict[str, ValueType],

) -> List[Tuple[EntityKeyProto, Dict[str, ValueProto], datetime, Optional[datetime]]]:

# Avoid ChunkedArrays which guarantees `zero_copy_only` available.

if isinstance(table, pyarrow.Table):

table = table.to_batches()[0]

columns = [

(field.name, field.dtype.to_value_type()) for field in feature_view.features

] + list(join_keys.items())

# Convert columns that exist in the table

proto_values_by_column = _columns_to_proto_values(

table, columns, allow_missing=True

)

# Ensure join keys are included, creating null values if missing from table

for join_key, value_type in join_keys.items():

if join_key not in proto_values_by_column:

if join_key in table.column_names:

proto_values_by_column[join_key] = python_values_to_proto_values(

table.column(join_key).to_numpy(zero_copy_only=False), value_type

)

else:

# Create null proto values directly (no need to build a PyArrow array)

proto_values_by_column[join_key] = python_values_to_proto_values(

[None] * table.num_rows, value_type

)

# Cache column names set to avoid recreating list in loop

column_names = {c[0] for c in columns}

# Adding On Demand Features that are missing from proto_values

for feature in feature_view.features:

if feature.name in column_names and feature.name not in proto_values_by_column:

# Create null proto values directly (more efficient than building PyArrow array)

proto_values_by_column[feature.name] = python_values_to_proto_values(

[None] * table.num_rows, feature.dtype.to_value_type()

)

entity_keys = _build_entity_keys(table.num_rows, join_keys, proto_values_by_column)

# Serialize the features per row

feature_dict = {

feature.name: proto_values_by_column[feature.name]

for feature in feature_view.features

if feature.name in proto_values_by_column

}

if feature_view.write_to_online_store:

table_columns = {col.name for col in table.schema}

for feature in feature_view.schema:

if feature.name not in feature_dict and feature.name in table_columns:

feature_dict[feature.name] = proto_values_by_column[feature.name]

features = [dict(zip(feature_dict, vars)) for vars in zip(*feature_dict.values())]

# We need to artificially add event_timestamps and created_timestamps

now = _utc_now()

event_timestamps = [

_coerce_datetime(pd.Timestamp(now)) for _ in range(table.num_rows)

]

# setting them equivalent

created_timestamps = event_timestamps

return list(zip(entity_keys, features, event_timestamps, created_timestamps))

def _validate_entity_values(join_key_values: Dict[str, List[ValueProto]]):

set_of_row_lengths = {len(v) for v in join_key_values.values()}

if len(set_of_row_lengths) > 1:

raise ValueError("All entity rows must have the same columns.")

return set_of_row_lengths.pop()

def _validate_feature_refs(feature_refs: List[str], full_feature_names: bool = False):

"""

Validates that there are no collisions among the feature references.

Args:

feature_refs: List of feature references to validate. Feature references must have format

"feature_view:feature", e.g. "customer_fv:daily_transactions".

full_feature_names: If True, the full feature references are compared for collisions; if False,

only the feature names are compared.

Raises:

FeatureNameCollisionError: There is a collision among the feature references.

"""

collided_feature_refs = []

if full_feature_names:

collided_feature_refs = [

ref for ref, occurrences in Counter(feature_refs).items() if occurrences > 1

]

else:

feature_names = [_parse_feature_ref(ref)[2] for ref in feature_refs]

collided_feature_names = [

ref

for ref, occurrences in Counter(feature_names).items()

if occurrences > 1

]

for feature_name in collided_feature_names:

collided_feature_refs.extend(

[ref for ref in feature_refs if ref.endswith(":" + feature_name)]

)

if len(collided_feature_refs) > 0:

raise FeatureNameCollisionError(collided_feature_refs, full_feature_names)

def _group_feature_refs(

features: List[str],

all_feature_views: List["FeatureView"],

all_on_demand_feature_views: List["OnDemandFeatureView"],

) -> Tuple[

List[Tuple[Union["FeatureView", "OnDemandFeatureView"], List[str]]],

List[Tuple["OnDemandFeatureView", List[str]]],

]:

"""Get list of feature views and corresponding feature names based on feature references"""

# view name to view proto

view_index: Dict[str, Union["FeatureView", "OnDemandFeatureView"]] = {

view.projection.name_to_use(): view for view in all_feature_views

}

# on demand view to on demand view proto

on_demand_view_index: Dict[str, "OnDemandFeatureView"] = {}

for view in all_on_demand_feature_views:

if view.projection and not view.write_to_online_store:

on_demand_view_index[view.projection.name_to_use()] = view

elif view.projection and view.write_to_online_store:

# we insert the ODFV view to FVs for ones that are written to the online store

view_index[view.projection.name_to_use()] = view

# view name to feature names

views_features = defaultdict(set)

# on demand view name to feature names

on_demand_view_features = defaultdict(set)

for ref in features:

fv_name, version_num, feat_name = _parse_feature_ref(ref)

# Build the key that matches projection.name_to_use()

if version_num is not None:

view_name = f"{fv_name}@v{version_num}"

else:

view_name = fv_name

if view_name in view_index:

if hasattr(view_index[view_name], "write_to_online_store"):

tmp_feat_name = [

f for f in view_index[view_name].schema if f.name == feat_name

]

if len(tmp_feat_name) > 0:

feat_name = tmp_feat_name[0].name

else:

view_index[view_name].projection.get_feature(

feat_name

) # For validation

views_features[view_name].add(feat_name)

elif view_name in on_demand_view_index:

on_demand_view_index[view_name].projection.get_feature(

feat_name

) # For validation

on_demand_view_features[view_name].add(feat_name)

# Let's also add in any FV Feature dependencies here.

for input_fv_projection in on_demand_view_index[

view_name

].source_feature_view_projections.values():

for input_feat in input_fv_projection.features:

views_features[input_fv_projection.name].add(input_feat.name)

else:

raise FeatureViewNotFoundException(view_name)

fvs_result: List[Tuple[Union["FeatureView", "OnDemandFeatureView"], List[str]]] = []

odfvs_result: List[Tuple["OnDemandFeatureView", List[str]]] = []

for view_name, feature_names in views_features.items():

fvs_result.append((view_index[view_name], list(feature_names)))

for view_name, feature_names in on_demand_view_features.items():

odfvs_result.append((on_demand_view_index[view_name], list(feature_names)))

return fvs_result, odfvs_result

def _apply_aggregations_to_response(

response_data: Union[pyarrow.Table, Dict[str, List[Any]]],

aggregations,

group_keys: Optional[List[str]],

mode: str,

) -> Union[pyarrow.Table, Dict[str, List[Any]]]:

"""

Apply aggregations using PandasBackend.

Args:

response_data: Either a pyarrow.Table or dict of lists containing the data

aggregations: List of Aggregation objects to apply

group_keys: List of column names to group by (optional)

mode: Transformation mode ("python", "pandas", or "substrait")

Returns:

Aggregated data in the same format as input

TODO: Consider refactoring to support backends other than pandas in the future.

"""

if not aggregations:

return response_data

backend = PandasBackend()

# Convert to pandas DataFrame

if isinstance(response_data, dict):

df = pd.DataFrame(response_data)

else: # pyarrow.Table

df = backend.from_arrow(response_data)

if df.empty:

return response_data

# Convert aggregations to agg_ops format

agg_ops = aggregation_specs_to_agg_ops(

aggregations,

time_window_unsupported_error_message=(

"Time window aggregation is not supported in online serving."

),

)

# Apply aggregations using PandasBackend

if group_keys:

result_df = backend.groupby_agg(df, group_keys, agg_ops)

else:

# No grouping - aggregate over entire dataset

result_df = backend.groupby_agg(df, [], agg_ops)

# Convert back to original format

if mode == "python":

return {col: result_df[col].tolist() for col in result_df.columns}

else: # pandas or substrait

return backend.to_arrow(result_df)

def _augment_response_with_on_demand_transforms(

online_features_response: GetOnlineFeaturesResponse,

feature_refs: List[str],

requested_on_demand_feature_views: List["OnDemandFeatureView"],

full_feature_names: bool,

feature_types: Optional[Dict[str, "ValueType"]] = None,

):

"""Computes on demand feature values and adds them to the result rows.

Assumes that 'online_features_response' already contains the necessary request data and input feature

views for the on demand feature views. Unneeded feature values such as request data and

unrequested input feature views will be removed from 'online_features_response'.

Args:

online_features_response: Protobuf object to populate

feature_refs: List of all feature references to be returned.

requested_on_demand_feature_views: List of all odfvs that have been requested.

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").

"""

from feast.online_response import OnlineResponse

requested_odfv_map = {odfv.name: odfv for odfv in requested_on_demand_feature_views}

requested_odfv_feature_names = requested_odfv_map.keys()

odfv_feature_refs = defaultdict(list)

for feature_ref in feature_refs:

view_name, _, feature_name = _parse_feature_ref(feature_ref)

if view_name in requested_odfv_feature_names:

odfv_feature_refs[view_name].append(

f"{requested_odfv_map[view_name].projection.name_to_use()}__{feature_name}"

if full_feature_names

else feature_name

)

initial_response = OnlineResponse(

online_features_response, feature_types=feature_types

)

initial_response_arrow: Optional[pyarrow.Table] = None

initial_response_dict: Optional[Dict[str, List[Any]]] = None

def _is_metrics_active():

try:

from feast.metrics import _config

return _config.online_features

except Exception:

return False

_metrics_active = _is_metrics_active()

# Apply on demand transformations and augment the result rows

odfv_result_names = set()

for odfv_name, _feature_refs in odfv_feature_refs.items():

odfv = requested_odfv_map[odfv_name]

if not odfv.write_to_online_store:

_should_track = _metrics_active and getattr(odfv, "track_metrics", False)

if _should_track:

import time as _time

_transform_start = _time.monotonic()

# Apply aggregations if configured.

if odfv.aggregations:

if odfv.mode == "python":

if initial_response_dict is None:

initial_response_dict = initial_response.to_dict()

initial_response_dict = _apply_aggregations_to_response(

initial_response_dict,

odfv.aggregations,

odfv.entities,

odfv.mode,

)

elif odfv.mode in {"pandas", "substrait"}:

if initial_response_arrow is None:

initial_response_arrow = initial_response.to_arrow()

initial_response_arrow = _apply_aggregations_to_response(

initial_response_arrow,

odfv.aggregations,

odfv.entities,

odfv.mode,

)

continue

# Apply transformation. Note: aggregations and transformation configs are mutually exclusive

# TODO: Fix to make it work for having both aggregation and transformation

# ticket: https://github.com/feast-dev/feast/issues/5689

# A UDF should only get features from the sources it declared. Drop

# features that belong to other feature views so it can't read them by

# accident (#6158). Join keys, request data and the declared features

# stay. substrait filters on its own, so leave it alone.

declared_source_names = {

projection.name

for projection in odfv.source_feature_view_projections.values()

}

undeclared_columns: set[str] = set()

# features the caller asked for that aren't one of our sources

for feature_ref in feature_refs:

ref_view_name, _, ref_feature_name = _parse_feature_ref(feature_ref)

if ref_view_name in requested_odfv_feature_names:

continue

if ref_view_name not in declared_source_names:

undeclared_columns.add(ref_feature_name)

undeclared_columns.add(f"{ref_view_name}__{ref_feature_name}")

# features that are only in the response because another requested

# ODFV needs them as input

for other_odfv in requested_on_demand_feature_views:

for projection in other_odfv.source_feature_view_projections.values():

if projection.name in declared_source_names:

continue

for feature in projection.features:

undeclared_columns.add(feature.name)

undeclared_columns.add(f"{projection.name}__{feature.name}")

if odfv.mode == "python":

if initial_response_dict is None:

initial_response_dict = initial_response.to_dict()

odfv_input_dict = {

k: v

for k, v in initial_response_dict.items()

if k not in undeclared_columns

}

transformed_features_dict: Dict[str, List[Any]] = odfv.transform_dict(

odfv_input_dict

)

elif odfv.mode in {"pandas", "substrait"}:

if initial_response_arrow is None:

initial_response_arrow = initial_response.to_arrow()

odfv_input_arrow = initial_response_arrow

if odfv.mode == "pandas":

odfv_input_arrow = initial_response_arrow.select(

[

c

for c in initial_response_arrow.column_names

if c not in undeclared_columns

]

)

transformed_features_arrow = odfv.transform_arrow(

odfv_input_arrow, full_feature_names

)