def get_historical_features(
config: RepoConfig,
feature_views: List[FeatureView],
feature_refs: List[str],
entity_df: Union[pd.DataFrame, str],
registry: Registry,
project: str,
full_feature_names: bool = False,
) -> RetrievalJob:
if not isinstance(entity_df, pd.DataFrame):
raise ValueError(
f"Please provide an entity_df of type {type(pd.DataFrame)} instead of type {type(entity_df)}"
)
entity_df_event_timestamp_col = DEFAULT_ENTITY_DF_EVENT_TIMESTAMP_COL # local modifiable copy of global variable
if entity_df_event_timestamp_col not in entity_df.columns:
datetime_columns = entity_df.select_dtypes(
include=["datetime", "datetimetz"]).columns
if len(datetime_columns) == 1:
print(
f"Using {datetime_columns[0]} as the event timestamp. To specify a column explicitly, please name it {DEFAULT_ENTITY_DF_EVENT_TIMESTAMP_COL}."
)
entity_df_event_timestamp_col = datetime_columns[0]
else:
raise ValueError(
f"Please provide an entity_df with a column named {DEFAULT_ENTITY_DF_EVENT_TIMESTAMP_COL} representing the time of events."
)
(
feature_views_to_features,
on_demand_feature_views_to_features,
) = _get_requested_feature_views_to_features_dict(
feature_refs,
feature_views,
registry.list_on_demand_feature_views(config.project),
)
# Create lazy function that is only called from the RetrievalJob object
def evaluate_historical_retrieval():
# Make sure all event timestamp fields are tz-aware. We default tz-naive fields to UTC
entity_df[entity_df_event_timestamp_col] = entity_df[
entity_df_event_timestamp_col].apply(
lambda x: x
if x.tzinfo is not None else x.replace(tzinfo=pytz.utc))
# Create a copy of entity_df to prevent modifying the original
entity_df_with_features = entity_df.copy()
# Convert event timestamp column to datetime and normalize time zone to UTC
# This is necessary to avoid issues with pd.merge_asof
entity_df_with_features[
entity_df_event_timestamp_col] = pd.to_datetime(
entity_df_with_features[entity_df_event_timestamp_col],
utc=True)
# Sort event timestamp values
entity_df_with_features = entity_df_with_features.sort_values(
entity_df_event_timestamp_col)
# Load feature view data from sources and join them incrementally
for feature_view, features in feature_views_to_features.items():
event_timestamp_column = (
feature_view.batch_source.event_timestamp_column)
created_timestamp_column = (
feature_view.batch_source.created_timestamp_column)
# Read offline parquet data in pyarrow format.
filesystem, path = FileSource.create_filesystem_and_path(
feature_view.batch_source.path,
feature_view.batch_source.file_options.
s3_endpoint_override,
)
table = pyarrow.parquet.read_table(path, filesystem=filesystem)
# Rename columns by the field mapping dictionary if it exists
if feature_view.batch_source.field_mapping is not None:
table = _run_field_mapping(
table, feature_view.batch_source.field_mapping)
# Rename entity columns by the join_key_map dictionary if it exists
if feature_view.projection.join_key_map:
table = _run_field_mapping(
table, feature_view.projection.join_key_map)
# Convert pyarrow table to pandas dataframe. Note, if the underlying data has missing values,
# pandas will convert those values to np.nan if the dtypes are numerical (floats, ints, etc.) or boolean
# If the dtype is 'object', then missing values are inferred as python `None`s.
# More details at:
# https://pandas.pydata.org/pandas-docs/stable/user_guide/missing_data.html#values-considered-missing
df_to_join = table.to_pandas()
# Make sure all timestamp fields are tz-aware. We default tz-naive fields to UTC
df_to_join[event_timestamp_column] = df_to_join[
event_timestamp_column].apply(lambda x: x
if x.tzinfo is not None else
x.replace(tzinfo=pytz.utc))
if created_timestamp_column:
df_to_join[created_timestamp_column] = df_to_join[
created_timestamp_column].apply(
lambda x: x if x.tzinfo is not None else x.replace(
tzinfo=pytz.utc))
# Sort dataframe by the event timestamp column
df_to_join = df_to_join.sort_values(event_timestamp_column)
# Build a list of all the features we should select from this source
feature_names = []
for feature in features:
# Modify the separator for feature refs in column names to double underscore. We are using
# double underscore as separator for consistency with other databases like BigQuery,
# where there are very few characters available for use as separators
if full_feature_names:
formatted_feature_name = (
f"{feature_view.projection.name_to_use()}__{feature}"
)
else:
formatted_feature_name = feature
# Add the feature name to the list of columns
feature_names.append(formatted_feature_name)
# Ensure that the source dataframe feature column includes the feature view name as a prefix
df_to_join.rename(
columns={feature: formatted_feature_name},
inplace=True,
)
# Build a list of entity columns to join on (from the right table)
join_keys = []
for entity_name in feature_view.entities:
entity = registry.get_entity(entity_name, project)
join_key = feature_view.projection.join_key_map.get(
entity.join_key, entity.join_key)
join_keys.append(join_key)
right_entity_columns = join_keys
right_entity_key_columns = [event_timestamp_column
] + right_entity_columns
# Remove all duplicate entity keys (using created timestamp)
right_entity_key_sort_columns = right_entity_key_columns
if created_timestamp_column:
# If created_timestamp is available, use it to dedupe deterministically
right_entity_key_sort_columns = right_entity_key_sort_columns + [
created_timestamp_column
]
df_to_join.sort_values(by=right_entity_key_sort_columns,
inplace=True)
df_to_join.drop_duplicates(
right_entity_key_sort_columns,
keep="last",
ignore_index=True,
inplace=True,
)
# Select only the columns we need to join from the feature dataframe
df_to_join = df_to_join[right_entity_key_columns +
feature_names]
# Do point in-time-join between entity_df and feature dataframe
entity_df_with_features = pd.merge_asof(
entity_df_with_features,
df_to_join,
left_on=entity_df_event_timestamp_col,
right_on=event_timestamp_column,
by=right_entity_columns or None,
tolerance=feature_view.ttl,
)
# Remove right (feature table/view) event_timestamp column.
if event_timestamp_column != entity_df_event_timestamp_col:
entity_df_with_features.drop(
columns=[event_timestamp_column], inplace=True)
# Ensure that we delete dataframes to free up memory
del df_to_join
# Move "event_timestamp" column to front
current_cols = entity_df_with_features.columns.tolist()
current_cols.remove(entity_df_event_timestamp_col)
entity_df_with_features = entity_df_with_features[
[entity_df_event_timestamp_col] + current_cols]
return entity_df_with_features
job = FileRetrievalJob(
evaluation_function=evaluate_historical_retrieval,
full_feature_names=full_feature_names,
on_demand_feature_views=OnDemandFeatureView.get_requested_odfvs(
feature_refs, project, registry),
)
return job
def get_historical_features(
config: RepoConfig,
feature_views: List[FeatureView],
feature_refs: List[str],
entity_df: Union[pd.DataFrame, str],
registry: Registry,
project: str,
full_feature_names: bool = False,
) -> RetrievalJob:
assert isinstance(config.offline_store, SnowflakeOfflineStoreConfig)
snowflake_conn = get_snowflake_conn(config.offline_store)
entity_schema = _get_entity_schema(entity_df, snowflake_conn, config)
entity_df_event_timestamp_col = offline_utils.infer_event_timestamp_from_entity_df(
entity_schema)
entity_df_event_timestamp_range = _get_entity_df_event_timestamp_range(
entity_df,
entity_df_event_timestamp_col,
snowflake_conn,
)
@contextlib.contextmanager
def query_generator() -> Iterator[str]:
table_name = offline_utils.get_temp_entity_table_name()
_upload_entity_df(entity_df, snowflake_conn, config, table_name)
expected_join_keys = offline_utils.get_expected_join_keys(
project, feature_views, registry)
offline_utils.assert_expected_columns_in_entity_df(
entity_schema, expected_join_keys,
entity_df_event_timestamp_col)
# Build a query context containing all information required to template the Snowflake SQL query
query_context = offline_utils.get_feature_view_query_context(
feature_refs,
feature_views,
registry,
project,
entity_df_event_timestamp_range,
)
query_context = _fix_entity_selections_identifiers(query_context)
# Generate the Snowflake SQL query from the query context
query = offline_utils.build_point_in_time_query(
query_context,
left_table_query_string=table_name,
entity_df_event_timestamp_col=entity_df_event_timestamp_col,
entity_df_columns=entity_schema.keys(),
query_template=MULTIPLE_FEATURE_VIEW_POINT_IN_TIME_JOIN,
full_feature_names=full_feature_names,
)
yield query
return SnowflakeRetrievalJob(
query=query_generator,
snowflake_conn=snowflake_conn,
config=config,
full_feature_names=full_feature_names,
on_demand_feature_views=OnDemandFeatureView.get_requested_odfvs(
feature_refs, project, registry),
metadata=RetrievalMetadata(
features=feature_refs,
keys=list(entity_schema.keys() -
{entity_df_event_timestamp_col}),
min_event_timestamp=entity_df_event_timestamp_range[0],
max_event_timestamp=entity_df_event_timestamp_range[1],
),
)
def get_historical_features(
config: RepoConfig,
feature_views: List[FeatureView],
feature_refs: List[str],
entity_df: Union[pandas.DataFrame, str],
registry: Registry,
project: str,
full_feature_names: bool = False,
) -> RetrievalJob:
assert isinstance(config.offline_store, SparkOfflineStoreConfig)
warnings.warn(
"The spark offline store is an experimental feature in alpha development. "
"Some functionality may still be unstable so functionality can change in the future.",
RuntimeWarning,
)
spark_session = get_spark_session_or_start_new_with_repoconfig(
store_config=config.offline_store)
tmp_entity_df_table_name = offline_utils.get_temp_entity_table_name()
entity_schema = _get_entity_schema(
spark_session=spark_session,
entity_df=entity_df,
)
event_timestamp_col = offline_utils.infer_event_timestamp_from_entity_df(
entity_schema=entity_schema, )
entity_df_event_timestamp_range = _get_entity_df_event_timestamp_range(
entity_df,
event_timestamp_col,
spark_session,
)
_upload_entity_df(
spark_session=spark_session,
table_name=tmp_entity_df_table_name,
entity_df=entity_df,
event_timestamp_col=event_timestamp_col,
)
expected_join_keys = offline_utils.get_expected_join_keys(
project=project, feature_views=feature_views, registry=registry)
offline_utils.assert_expected_columns_in_entity_df(
entity_schema=entity_schema,
join_keys=expected_join_keys,
entity_df_event_timestamp_col=event_timestamp_col,
)
query_context = offline_utils.get_feature_view_query_context(
feature_refs,
feature_views,
registry,
project,
entity_df_event_timestamp_range,
)
query = offline_utils.build_point_in_time_query(
feature_view_query_contexts=query_context,
left_table_query_string=tmp_entity_df_table_name,
entity_df_event_timestamp_col=event_timestamp_col,
entity_df_columns=entity_schema.keys(),
query_template=MULTIPLE_FEATURE_VIEW_POINT_IN_TIME_JOIN,
full_feature_names=full_feature_names,
)
return SparkRetrievalJob(
spark_session=spark_session,
query=query,
full_feature_names=full_feature_names,
on_demand_feature_views=OnDemandFeatureView.get_requested_odfvs(
feature_refs, project, registry),
metadata=RetrievalMetadata(
features=feature_refs,
keys=list(set(entity_schema.keys()) - {event_timestamp_col}),
min_event_timestamp=entity_df_event_timestamp_range[0],
max_event_timestamp=entity_df_event_timestamp_range[1],
),
)
def get_historical_features(
config: RepoConfig,
feature_views: List[FeatureView],
feature_refs: List[str],
entity_df: Union[pd.DataFrame, str],
registry: Registry,
project: str,
full_feature_names: bool = False,
) -> RetrievalJob:
assert isinstance(config.offline_store, RedshiftOfflineStoreConfig)
redshift_client = aws_utils.get_redshift_data_client(
config.offline_store.region
)
s3_resource = aws_utils.get_s3_resource(config.offline_store.region)
@contextlib.contextmanager
def query_generator() -> Iterator[str]:
table_name = offline_utils.get_temp_entity_table_name()
entity_schema = _upload_entity_df_and_get_entity_schema(
entity_df, redshift_client, config, s3_resource, table_name
)
entity_df_event_timestamp_col = offline_utils.infer_event_timestamp_from_entity_df(
entity_schema
)
expected_join_keys = offline_utils.get_expected_join_keys(
project, feature_views, registry
)
offline_utils.assert_expected_columns_in_entity_df(
entity_schema, expected_join_keys, entity_df_event_timestamp_col
)
# Build a query context containing all information required to template the Redshift SQL query
query_context = offline_utils.get_feature_view_query_context(
feature_refs, feature_views, registry, project,
)
# Generate the Redshift SQL query from the query context
query = offline_utils.build_point_in_time_query(
query_context,
left_table_query_string=table_name,
entity_df_event_timestamp_col=entity_df_event_timestamp_col,
query_template=MULTIPLE_FEATURE_VIEW_POINT_IN_TIME_JOIN,
full_feature_names=full_feature_names,
)
yield query
# Clean up the uploaded Redshift table
aws_utils.execute_redshift_statement(
redshift_client,
config.offline_store.cluster_id,
config.offline_store.database,
config.offline_store.user,
f"DROP TABLE {table_name}",
)
return RedshiftRetrievalJob(
query=query_generator,
redshift_client=redshift_client,
s3_resource=s3_resource,
config=config,
full_feature_names=full_feature_names,
on_demand_feature_views=OnDemandFeatureView.get_requested_odfvs(
feature_refs, project, registry
),
drop_columns=["entity_timestamp"]
+ [
f"{feature_view.name}__entity_row_unique_id"
for feature_view in feature_views
],
)
def get_historical_features(
config: RepoConfig,
feature_views: List[FeatureView],
feature_refs: List[str],
entity_df: Union[pd.DataFrame, str],
registry: Registry,
project: str,
full_feature_names: bool = False,
) -> RetrievalJob:
if not isinstance(entity_df, pd.DataFrame) and not isinstance(
entity_df, dd.DataFrame):
raise ValueError(
f"Please provide an entity_df of type {type(pd.DataFrame)} instead of type {type(entity_df)}"
)
entity_df_event_timestamp_col = DEFAULT_ENTITY_DF_EVENT_TIMESTAMP_COL # local modifiable copy of global variable
if entity_df_event_timestamp_col not in entity_df.columns:
datetime_columns = entity_df.select_dtypes(
include=["datetime", "datetimetz"]).columns
if len(datetime_columns) == 1:
print(
f"Using {datetime_columns[0]} as the event timestamp. To specify a column explicitly, please name it {DEFAULT_ENTITY_DF_EVENT_TIMESTAMP_COL}."
)
entity_df_event_timestamp_col = datetime_columns[0]
else:
raise ValueError(
f"Please provide an entity_df with a column named {DEFAULT_ENTITY_DF_EVENT_TIMESTAMP_COL} representing the time of events."
)
(
feature_views_to_features,
on_demand_feature_views_to_features,
) = _get_requested_feature_views_to_features_dict(
feature_refs,
feature_views,
registry.list_on_demand_feature_views(config.project),
)
entity_df_event_timestamp_range = _get_entity_df_event_timestamp_range(
entity_df, entity_df_event_timestamp_col)
# Create lazy function that is only called from the RetrievalJob object
def evaluate_historical_retrieval():
# Create a copy of entity_df to prevent modifying the original
entity_df_with_features = entity_df.copy()
entity_df_event_timestamp_col_type = entity_df_with_features.dtypes[
entity_df_event_timestamp_col]
if (not hasattr(entity_df_event_timestamp_col_type, "tz")
or entity_df_event_timestamp_col_type.tz != pytz.UTC):
# Make sure all event timestamp fields are tz-aware. We default tz-naive fields to UTC
entity_df_with_features[
entity_df_event_timestamp_col] = entity_df_with_features[
entity_df_event_timestamp_col].apply(
lambda x: x if x.tzinfo is not None else x.replace(
tzinfo=pytz.utc))
# Convert event timestamp column to datetime and normalize time zone to UTC
# This is necessary to avoid issues with pd.merge_asof
if isinstance(entity_df_with_features, dd.DataFrame):
entity_df_with_features[
entity_df_event_timestamp_col] = dd.to_datetime(
entity_df_with_features[
entity_df_event_timestamp_col],
utc=True)
else:
entity_df_with_features[
entity_df_event_timestamp_col] = pd.to_datetime(
entity_df_with_features[
entity_df_event_timestamp_col],
utc=True)
# Sort event timestamp values
entity_df_with_features = entity_df_with_features.sort_values(
entity_df_event_timestamp_col)
join_keys = []
all_join_keys = []
# Load feature view data from sources and join them incrementally
for feature_view, features in feature_views_to_features.items():
event_timestamp_column = feature_view.batch_source.timestamp_field
created_timestamp_column = (
feature_view.batch_source.created_timestamp_column)
# Build a list of entity columns to join on (from the right table)
join_keys = []
for entity_name in feature_view.entities:
entity = registry.get_entity(entity_name, project)
join_key = feature_view.projection.join_key_map.get(
entity.join_key, entity.join_key)
join_keys.append(join_key)
right_entity_key_columns = [
event_timestamp_column,
created_timestamp_column,
] + join_keys
right_entity_key_columns = [
c for c in right_entity_key_columns if c
]
all_join_keys = list(set(all_join_keys + join_keys))
df_to_join = _read_datasource(feature_view.batch_source)
df_to_join, event_timestamp_column = _field_mapping(
df_to_join,
feature_view,
features,
right_entity_key_columns,
entity_df_event_timestamp_col,
event_timestamp_column,
full_feature_names,
)
df_to_join = _merge(entity_df_with_features, df_to_join,
join_keys)
df_to_join = _normalize_timestamp(df_to_join,
event_timestamp_column,
created_timestamp_column)
df_to_join = _filter_ttl(
df_to_join,
feature_view,
entity_df_event_timestamp_col,
event_timestamp_column,
)
df_to_join = _drop_duplicates(
df_to_join,
all_join_keys,
event_timestamp_column,
created_timestamp_column,
entity_df_event_timestamp_col,
)
entity_df_with_features = _drop_columns(
df_to_join, event_timestamp_column,
created_timestamp_column)
# Ensure that we delete dataframes to free up memory
del df_to_join
return entity_df_with_features.persist()
job = FileRetrievalJob(
evaluation_function=evaluate_historical_retrieval,
full_feature_names=full_feature_names,
on_demand_feature_views=OnDemandFeatureView.get_requested_odfvs(
feature_refs, project, registry),
metadata=RetrievalMetadata(
features=feature_refs,
keys=list(
set(entity_df.columns) - {entity_df_event_timestamp_col}),
min_event_timestamp=entity_df_event_timestamp_range[0],
max_event_timestamp=entity_df_event_timestamp_range[1],
),
)
return job
