marlinfs 0.0.2

Python Client for Marlin Feature Store

Documentation

See the API Docs.

Installing

pip install marlinfs

Usage

Login

marlin.login()

Batch Ingestion

transform_client = marlin.transform_client(namespace, name, version, entities)


@transform_client.process_function
def process():
    dep1 = transform_client.add_dependency('n1', 't1', 'v1', ['f1', 'f2'])
    # Reading by timestamp
    ingestion_time_read = dep1.read_by_ingestion_ts(1612140982, 1612150982)
    event_time_read = dep1.read_by_event_ts(1612140982, 1612150982)

    # Reading by date
    ingestion_date_read = dep1.read_by_ingestion_date("2021-01-02-03", "2021-01-02-04")
    event_date_read = dep1.read_by_event_date("2021-01-02-03", "2021-01-02-04")

    # To commit metadata and store data
    transform_client.commit()

    # assumption is df contains event_timestamp column in date in this format: 2021-01-02-03. 
    # To pass different date format change return statement to df, {'date_format': 'str_date', 'str_date_format_type': '<python date format>' e.g. %Y-%m-%d})
    # To pass event_timestamp in seconds change return statement to df, {'date_format': 'seconds'}
    return df 

Batch Serving

batch_serving_client = marlin.batch_training_client(namespace, name, version)


# batch_serving_client = marlin.batch_scoring_client(namespace, name, version)

@batch_serving_client.serving_function
def process():
    entity_df = None  # Some entity df

    dep1 = batch_serving_client.add_dependency('n1', 't1', 'v1', ['f1', 'f2'])
    dep2 = batch_serving_client.add_dependency('n2', 't1', 'v1', ['f1', 'f2'])

    entity_df = pd.DataFrame([
        [1, 1, 1, 1, "2021-01-02-03"],
        [1, 1, 1, 1, "2021-01-02-03"]
    ], columns=['A', 'B', 'C', 'D', 'target_timestamp'])
    dep1.point_in_time_join_by_date(entity_df)
    dep1.point_in_time_join_across_inputs_by_date(entity_df, [dep2])

    entity_df = pd.DataFrame([
        [1, 1, 1, 1, 1612140982],
        [1, 1, 1, 1, 1612140982]
    ], columns=['A', 'B', 'C', 'D', 'target_timestamp'])
    dep1.point_in_time_join_by_ts(entity_df)
    dep1.point_in_time_join_across_inputs_by_ts(entity_df, [dep2])

    # To commit metadata
    batch_serving_client.commit()

Exploration Client

exploration_client = marlin.exploration_client()
tf1 = exploration_client.get_transform('n1', 't1', 'v1')
tf2 = exploration_client.get_transform('n2', 't2', 'v1')

entity_df = pd.DataFrame([
    [1, 1, 1, 1, "2021-01-02-03"],
    [1, 1, 1, 1, "2021-01-02-03"]
], columns=['A', 'B', 'C', 'D', 'target_timestamp'])

tf1.point_in_time_join_by_date(entity_df)
tf1.point_in_time_join_across_inputs_by_date(entity_df, [tf2])

entity_df = pd.DataFrame([
    [1, 1, 1, 1, 1612140982],
    [1, 1, 1, 1, 1612140982]
], columns=['A', 'B', 'C', 'D', 'target_timestamp'])
tf1.point_in_time_join_by_ts(entity_df)
tf1.point_in_time_join_across_inputs_by_ts(entity_df, [tf2])

# Reading by timestamp
ingestion_time_read = tf1.read_by_ingestion_ts(1612140982, 1612150982)
event_time_read = tf1.read_by_event_ts(1612140982, 1612150982)

# Reading by date
ingestion_date_read = tf1.read_by_ingestion_date("2021-01-02-03", "2021-01-02-04")
event_date_read = tf1.read_by_event_date("2021-01-02-03", "2021-01-02-04")