def test_source(
self,
target_df_source,
target_df_table_reader,
spark_session,
):
# given
spark_client = SparkClient()
table_reader_id = "a_test_source"
table_reader_db = "db"
table_reader_table = "table_test_source"
create_temp_view(dataframe=target_df_table_reader,
name=table_reader_id)
create_db_and_table(
spark=spark_session,
table_reader_id=table_reader_id,
table_reader_db=table_reader_db,
table_reader_table=table_reader_table,
)
file_reader_id = "b_test_source"
data_sample_path = INPUT_PATH + "/data.json"
# when
source = Source(
readers=[
TableReader(
id=table_reader_id,
database=table_reader_db,
table=table_reader_table,
),
FileReader(id=file_reader_id,
path=data_sample_path,
format="json"),
],
query=f"select a.*, b.feature2 " # noqa
f"from {table_reader_id} a " # noqa
f"inner join {file_reader_id} b on a.id = b.id ", # noqa
)
result_df = source.construct(client=spark_client)
target_df = target_df_source
# then
assert (compare_dataframes(
actual_df=result_df,
expected_df=target_df,
columns_sort=result_df.columns,
) is True)
def __init__(self):
super(FirstPipeline, self).__init__(
source=Source(
readers=[TableReader(id="t", database="db", table="table",)],
query=f"select * from t", # noqa
),
feature_set=FeatureSet(
name="first",
entity="entity",
description="description",
features=[
Feature(name="feature1", description="test", dtype=DataType.FLOAT,),
Feature(
name="feature2",
description="another test",
dtype=DataType.STRING,
),
],
keys=[
KeyFeature(
name="id", description="identifier", dtype=DataType.BIGINT,
)
],
timestamp=TimestampFeature(),
),
sink=Sink(
writers=[HistoricalFeatureStoreWriter(), OnlineFeatureStoreWriter()]
),
)
def __init__(self):
super(UserChargebacksPipeline, self).__init__(
source=Source(
readers=[
FileReader(
id="chargeback_events",
path="data/order_events/input.csv",
format="csv",
format_options={"header": True},
)
],
query=("""
select
cpf,
timestamp(chargeback_timestamp) as timestamp,
order_id
from
chargeback_events
where
chargeback_timestamp is not null
"""),
),
feature_set=AggregatedFeatureSet(
name="user_chargebacks",
entity="user",
description="Aggregates the total of chargebacks from users in "
"different time windows.",
keys=[
KeyFeature(
name="cpf",
description="User unique identifier, entity key.",
dtype=DataType.STRING,
)
],
timestamp=TimestampFeature(),
features=[
Feature(
name="cpf_chargebacks",
description=
"Total of chargebacks registered on user's CPF",
transformation=AggregatedTransform(functions=[
Function(functions.count, DataType.INTEGER)
]),
from_column="order_id",
),
],
).with_windows(
definitions=["3 days", "7 days", "30 days"]).add_post_hook(
ZeroFillHook()),
sink=Sink(writers=[
LocalHistoricalFSWriter(),
OnlineFeatureStoreWriter(
interval_mode=True,
check_schema_hook=NotCheckSchemaHook(),
debug_mode=True,
),
]),
)
def test_pipeline_with_hooks(self, spark_session):
# arrange
hook1 = AddHook(value=1)
spark_session.sql(
"select 1 as id, timestamp('2020-01-01') as timestamp, 0 as feature"
).createOrReplaceTempView("test")
target_df = spark_session.sql(
"select 1 as id, timestamp('2020-01-01') as timestamp, 6 as feature, 2020 "
"as year, 1 as month, 1 as day")
historical_writer = HistoricalFeatureStoreWriter(debug_mode=True)
test_pipeline = FeatureSetPipeline(
source=Source(
readers=[
TableReader(
id="reader",
table="test",
).add_post_hook(hook1)
],
query="select * from reader",
).add_post_hook(hook1),
feature_set=FeatureSet(
name="feature_set",
entity="entity",
description="description",
features=[
Feature(
name="feature",
description="test",
transformation=SQLExpressionTransform(
expression="feature + 1"),
dtype=DataType.INTEGER,
),
],
keys=[
KeyFeature(
name="id",
description="The user's Main ID or device ID",
dtype=DataType.INTEGER,
)
],
timestamp=TimestampFeature(),
).add_pre_hook(hook1).add_post_hook(hook1),
sink=Sink(writers=[historical_writer], ).add_pre_hook(hook1),
)
# act
test_pipeline.run()
output_df = spark_session.table(
"historical_feature_store__feature_set")
# assert
output_df.show()
assert_dataframe_equality(output_df, target_df)
def test_construct(self, mocker, target_df):
# given
spark_client = SparkClient()
reader_id = "a_source"
reader = mocker.stub(reader_id)
reader.build = mocker.stub("build")
reader.build.side_effect = target_df.createOrReplaceTempView(reader_id)
# when
source_selector = Source(
readers=[reader],
query=f"select * from {reader_id}", # noqa
)
result_df = source_selector.construct(spark_client)
assert result_df.collect() == target_df.collect()
def feature_set_pipeline(
spark_context, spark_session,
):
feature_set_pipeline = FeatureSetPipeline(
source=Source(
readers=[
TableReader(id="b_source", table="b_table",).with_incremental_strategy(
incremental_strategy=IncrementalStrategy(column="timestamp")
),
],
query=f"select * from b_source ", # noqa
),
feature_set=FeatureSet(
name="feature_set",
entity="entity",
description="description",
features=[
Feature(
name="feature",
description="test",
transformation=SparkFunctionTransform(
functions=[
Function(F.avg, DataType.FLOAT),
Function(F.stddev_pop, DataType.FLOAT),
],
).with_window(
partition_by="id",
order_by=TIMESTAMP_COLUMN,
mode="fixed_windows",
window_definition=["1 day"],
),
),
],
keys=[
KeyFeature(
name="id",
description="The user's Main ID or device ID",
dtype=DataType.INTEGER,
)
],
timestamp=TimestampFeature(),
),
sink=Sink(writers=[HistoricalFeatureStoreWriter(debug_mode=True)]),
)
return feature_set_pipeline
def test_feature_set_args(self):
# arrange and act
out_columns = [
"user_id",
"timestamp",
"listing_page_viewed__rent_per_month__avg_over_7_days_fixed_windows",
"listing_page_viewed__rent_per_month__avg_over_2_weeks_fixed_windows",
"listing_page_viewed__rent_per_month__stddev_pop_over_7_days_fixed_windows",
"listing_page_viewed__rent_per_month__"
"stddev_pop_over_2_weeks_fixed_windows",
# noqa
]
pipeline = FeatureSetPipeline(
source=Source(
readers=[
TableReader(
id="source_a",
database="db",
table="table",
),
FileReader(
id="source_b",
path="path",
format="parquet",
),
],
query="select a.*, b.specific_feature "
"from source_a left join source_b on a.id=b.id",
),
feature_set=FeatureSet(
name="feature_set",
entity="entity",
description="description",
keys=[
KeyFeature(
name="user_id",
description="The user's Main ID or device ID",
dtype=DataType.INTEGER,
)
],
timestamp=TimestampFeature(from_column="ts"),
features=[
Feature(
name="listing_page_viewed__rent_per_month",
description="Average of something.",
transformation=SparkFunctionTransform(functions=[
Function(functions.avg, DataType.FLOAT),
Function(functions.stddev_pop, DataType.FLOAT),
], ).with_window(
partition_by="user_id",
order_by=TIMESTAMP_COLUMN,
window_definition=["7 days", "2 weeks"],
mode="fixed_windows",
),
),
],
),
sink=Sink(writers=[
HistoricalFeatureStoreWriter(db_config=None),
OnlineFeatureStoreWriter(db_config=None),
], ),
)
assert isinstance(pipeline.spark_client, SparkClient)
assert len(pipeline.source.readers) == 2
assert all(
isinstance(reader, Reader) for reader in pipeline.source.readers)
assert isinstance(pipeline.source.query, str)
assert pipeline.feature_set.name == "feature_set"
assert pipeline.feature_set.entity == "entity"
assert pipeline.feature_set.description == "description"
assert isinstance(pipeline.feature_set.timestamp, TimestampFeature)
assert len(pipeline.feature_set.keys) == 1
assert all(
isinstance(k, KeyFeature) for k in pipeline.feature_set.keys)
assert len(pipeline.feature_set.features) == 1
assert all(
isinstance(feature, Feature)
for feature in pipeline.feature_set.features)
assert pipeline.feature_set.columns == out_columns
assert len(pipeline.sink.writers) == 2
assert all(
isinstance(writer, Writer) for writer in pipeline.sink.writers)
def test_feature_set_pipeline(
self,
mocked_df,
spark_session,
fixed_windows_output_feature_set_dataframe,
):
# arrange
table_reader_id = "a_source"
table_reader_table = "table"
table_reader_db = environment.get_variable(
"FEATURE_STORE_HISTORICAL_DATABASE")
create_temp_view(dataframe=mocked_df, name=table_reader_id)
create_db_and_table(
spark=spark_session,
table_reader_id=table_reader_id,
table_reader_db=table_reader_db,
table_reader_table=table_reader_table,
)
dbconfig = Mock()
dbconfig.mode = "overwrite"
dbconfig.format_ = "parquet"
dbconfig.get_options = Mock(
return_value={"path": "test_folder/historical/entity/feature_set"})
historical_writer = HistoricalFeatureStoreWriter(db_config=dbconfig)
# act
test_pipeline = FeatureSetPipeline(
source=Source(
readers=[
TableReader(
id=table_reader_id,
database=table_reader_db,
table=table_reader_table,
),
],
query=f"select * from {table_reader_id} ", # noqa
),
feature_set=FeatureSet(
name="feature_set",
entity="entity",
description="description",
features=[
Feature(
name="feature1",
description="test",
transformation=SparkFunctionTransform(functions=[
Function(F.avg, DataType.FLOAT),
Function(F.stddev_pop, DataType.FLOAT),
], ).with_window(
partition_by="id",
order_by=TIMESTAMP_COLUMN,
mode="fixed_windows",
window_definition=["2 minutes", "15 minutes"],
),
),
Feature(
name="divided_feature",
description="unit test",
dtype=DataType.FLOAT,
transformation=CustomTransform(
transformer=divide,
column1="feature1",
column2="feature2",
),
),
],
keys=[
KeyFeature(
name="id",
description="The user's Main ID or device ID",
dtype=DataType.INTEGER,
)
],
timestamp=TimestampFeature(),
),
sink=Sink(writers=[historical_writer]),
)
test_pipeline.run()
# assert
path = dbconfig.get_options("historical/entity/feature_set").get(
"path")
df = spark_session.read.parquet(path).orderBy(TIMESTAMP_COLUMN)
target_df = fixed_windows_output_feature_set_dataframe.orderBy(
test_pipeline.feature_set.timestamp_column)
# assert
assert_dataframe_equality(df, target_df)
# tear down
shutil.rmtree("test_folder")
def test_pipeline_interval_run(self, mocked_date_df,
pipeline_interval_run_target_dfs,
spark_session):
"""Testing pipeline's idempotent interval run feature.
Source data:
+-------+---+-------------------+-------------------+
|feature| id| ts| timestamp|
+-------+---+-------------------+-------------------+
| 200| 1|2016-04-11 11:31:11|2016-04-11 11:31:11|
| 300| 1|2016-04-12 11:44:12|2016-04-12 11:44:12|
| 400| 1|2016-04-13 11:46:24|2016-04-13 11:46:24|
| 500| 1|2016-04-14 12:03:21|2016-04-14 12:03:21|
+-------+---+-------------------+-------------------+
The test executes 3 runs for different time intervals. The input data has 4 data
points: 2016-04-11, 2016-04-12, 2016-04-13 and 2016-04-14. The following run
specifications are:
1) Interval: from 2016-04-11 to 2016-04-13
Target table result:
+---+-------+---+-----+------+-------------------+----+
|day|feature| id|month|run_id| timestamp|year|
+---+-------+---+-----+------+-------------------+----+
| 11| 200| 1| 4| 1|2016-04-11 11:31:11|2016|
| 12| 300| 1| 4| 1|2016-04-12 11:44:12|2016|
| 13| 400| 1| 4| 1|2016-04-13 11:46:24|2016|
+---+-------+---+-----+------+-------------------+----+
2) Interval: only 2016-04-14.
Target table result:
+---+-------+---+-----+------+-------------------+----+
|day|feature| id|month|run_id| timestamp|year|
+---+-------+---+-----+------+-------------------+----+
| 11| 200| 1| 4| 1|2016-04-11 11:31:11|2016|
| 12| 300| 1| 4| 1|2016-04-12 11:44:12|2016|
| 13| 400| 1| 4| 1|2016-04-13 11:46:24|2016|
| 14| 500| 1| 4| 2|2016-04-14 12:03:21|2016|
+---+-------+---+-----+------+-------------------+----+
3) Interval: only 2016-04-11.
Target table result:
+---+-------+---+-----+------+-------------------+----+
|day|feature| id|month|run_id| timestamp|year|
+---+-------+---+-----+------+-------------------+----+
| 11| 200| 1| 4| 3|2016-04-11 11:31:11|2016|
| 12| 300| 1| 4| 1|2016-04-12 11:44:12|2016|
| 13| 400| 1| 4| 1|2016-04-13 11:46:24|2016|
| 14| 500| 1| 4| 2|2016-04-14 12:03:21|2016|
+---+-------+---+-----+------+-------------------+----+
"""
# arrange
create_temp_view(dataframe=mocked_date_df, name="input_data")
db = environment.get_variable("FEATURE_STORE_HISTORICAL_DATABASE")
path = "test_folder/historical/entity/feature_set"
spark_session.conf.set("spark.sql.sources.partitionOverwriteMode",
"dynamic")
spark_session.sql(f"create database if not exists {db}")
spark_session.sql(
f"create table if not exists {db}.feature_set_interval "
f"(id int, timestamp timestamp, feature int, "
f"run_id int, year int, month int, day int);")
dbconfig = MetastoreConfig()
dbconfig.get_options = Mock(return_value={
"mode": "overwrite",
"format_": "parquet",
"path": path
})
historical_writer = HistoricalFeatureStoreWriter(db_config=dbconfig,
interval_mode=True)
first_run_hook = RunHook(id=1)
second_run_hook = RunHook(id=2)
third_run_hook = RunHook(id=3)
(
first_run_target_df,
second_run_target_df,
third_run_target_df,
) = pipeline_interval_run_target_dfs
test_pipeline = FeatureSetPipeline(
source=Source(
readers=[
TableReader(
id="id",
table="input_data",
).with_incremental_strategy(IncrementalStrategy("ts")),
],
query="select * from id ",
),
feature_set=FeatureSet(
name="feature_set_interval",
entity="entity",
description="",
keys=[
KeyFeature(
name="id",
description="",
dtype=DataType.INTEGER,
)
],
timestamp=TimestampFeature(from_column="ts"),
features=[
Feature(name="feature",
description="",
dtype=DataType.INTEGER),
Feature(name="run_id",
description="",
dtype=DataType.INTEGER),
],
),
sink=Sink([historical_writer], ),
)
# act and assert
dbconfig.get_path_with_partitions = Mock(return_value=[
"test_folder/historical/entity/feature_set/year=2016/month=4/day=11",
"test_folder/historical/entity/feature_set/year=2016/month=4/day=12",
"test_folder/historical/entity/feature_set/year=2016/month=4/day=13",
])
test_pipeline.feature_set.add_pre_hook(first_run_hook)
test_pipeline.run(end_date="2016-04-13", start_date="2016-04-11")
first_run_output_df = spark_session.read.parquet(path)
assert_dataframe_equality(first_run_output_df, first_run_target_df)
dbconfig.get_path_with_partitions = Mock(return_value=[
"test_folder/historical/entity/feature_set/year=2016/month=4/day=14",
])
test_pipeline.feature_set.add_pre_hook(second_run_hook)
test_pipeline.run_for_date("2016-04-14")
second_run_output_df = spark_session.read.parquet(path)
assert_dataframe_equality(second_run_output_df, second_run_target_df)
dbconfig.get_path_with_partitions = Mock(return_value=[
"test_folder/historical/entity/feature_set/year=2016/month=4/day=11",
])
test_pipeline.feature_set.add_pre_hook(third_run_hook)
test_pipeline.run_for_date("2016-04-11")
third_run_output_df = spark_session.read.parquet(path)
assert_dataframe_equality(third_run_output_df, third_run_target_df)
# tear down
shutil.rmtree("test_folder")
def __init__(self):
super(AwesomeDatasetPipeline, self).__init__(
source=Source(
readers=[
FileReader(
id="order_events",
path="data/order_events/input.csv",
format="csv",
format_options={"header": True},
),
FileReader(
id="user_chargebacks",
path="data/feature_store/historical/user/user_chargebacks",
format="parquet",
),
FileReader(
id="user_orders",
path="data/feature_store/historical/user/user_orders",
format="parquet",
),
],
query="""
with feature_sets_merge as(
select
user_orders.cpf,
user_orders.timestamp,
user_chargebacks.timestamp as chargeback_timestamp,
cpf_orders__count_over_3_days_rolling_windows,
cpf_orders__count_over_7_days_rolling_windows,
cpf_orders__count_over_30_days_rolling_windows,
cpf_chargebacks__count_over_3_days_rolling_windows,
cpf_chargebacks__count_over_7_days_rolling_windows,
cpf_chargebacks__count_over_30_days_rolling_windows,
row_number() over (
partition by (user_orders.cpf, user_orders.timestamp)
order by user_chargebacks.timestamp desc
) as rn
from
user_orders
left join user_chargebacks
on user_orders.cpf = user_chargebacks.cpf
and user_orders.timestamp >= user_chargebacks.timestamp
),
feature_sets_rn_filter as(
select
*
from
feature_sets_merge
where
rn = 1
),
orders_with_feature_sets as(
select
order_events.order_id,
timestamp(order_events.order_timestamp) as timestamp,
timestamp(order_events.chargeback_timestamp) as chargeback_timestamp,
order_events.cpf,
feature_sets_rn_filter.cpf_orders__count_over_3_days_rolling_windows,
feature_sets_rn_filter.cpf_orders__count_over_7_days_rolling_windows,
feature_sets_rn_filter.cpf_orders__count_over_30_days_rolling_windows,
feature_sets_rn_filter.cpf_chargebacks__count_over_3_days_rolling_windows,
feature_sets_rn_filter.cpf_chargebacks__count_over_7_days_rolling_windows,
feature_sets_rn_filter.cpf_chargebacks__count_over_30_days_rolling_windows,
row_number() over (
partition by (order_events.cpf, order_events.order_timestamp)
order by feature_sets_rn_filter.timestamp desc
) as rn
from
order_events
join feature_sets_rn_filter
on order_events.cpf = feature_sets_rn_filter.cpf
and timestamp(order_events.order_timestamp) >=
feature_sets_rn_filter.timestamp
)
select
order_id,
timestamp,
chargeback_timestamp,
cpf,
cpf_orders__count_over_3_days_rolling_windows,
cpf_orders__count_over_7_days_rolling_windows,
cpf_orders__count_over_30_days_rolling_windows,
coalesce(
cpf_chargebacks__count_over_3_days_rolling_windows,
0) as cpf_chargebacks__count_over_3_days_rolling_windows,
coalesce(
cpf_chargebacks__count_over_7_days_rolling_windows,
0) as cpf_chargebacks__count_over_7_days_rolling_windows,
coalesce(
cpf_chargebacks__count_over_30_days_rolling_windows,
0) as cpf_chargebacks__count_over_30_days_rolling_windows
from
orders_with_feature_sets
where
rn = 1
""",
),
feature_set=FeatureSet(
name="awesome_dataset",
entity="user",
description="Dataset enriching orders events with aggregated features "
"on total of orders and chargebacks by user.",
keys=[
KeyFeature(
name="order_id",
description="Orders unique identifier.",
dtype=DataType.STRING,
)
],
timestamp=TimestampFeature(),
features=[
Feature(
name="chargeback_timestamp",
description="Timestamp for the order creation.",
dtype=DataType.TIMESTAMP,
),
Feature(
name="cpf",
description="User unique identifier, user entity key.",
dtype=DataType.STRING,
),
Feature(
name="cpf_orders__count_over_3_days_rolling_windows",
description="Count of orders over 3 days rolling windows group "
"by user (identified by CPF)",
dtype=DataType.INTEGER,
),
Feature(
name="cpf_orders__count_over_7_days_rolling_windows",
description="Count of orders over 7 days rolling windows group "
"by user (identified by CPF)",
dtype=DataType.INTEGER,
),
Feature(
name="cpf_orders__count_over_30_days_rolling_windows",
description="Count of orders over 30 days rolling windows group"
" by user (identified by CPF)",
dtype=DataType.INTEGER,
),
Feature(
name="cpf_chargebacks__count_over_3_days_rolling_windows",
description="Count of chargebacks over 3 days rolling windows "
"group by user (identified by CPF)",
dtype=DataType.INTEGER,
),
Feature(
name="cpf_chargebacks__count_over_7_days_rolling_windows",
description="Count of chargebacks over 7 days rolling windows "
"group by user (identified by CPF)",
dtype=DataType.INTEGER,
),
Feature(
name="cpf_chargebacks__count_over_30_days_rolling_windows",
description="Count of chargebacks over 30 days rolling windows "
"group by user (identified by CPF)",
dtype=DataType.INTEGER,
),
],
),
sink=Sink(writers=[DatasetWriter()]),
)