Feature Monitoring

feathr_project/feathr/__init__.py

@@ -9,6 +9,7 @@
 from .definition.source import *
 from .definition.typed_key import *
 from .definition.materialization_settings import *
+from .definition.monitoring_settings import *
 from .definition.sink import *
 from .definition.query_feature_list import *
 from .definition.lookup_feature import *
@@ -19,7 +20,7 @@
 from .api.app.core.feathr_api_exception import *
 
 # skipped class as they are internal methods:
-# RepoDefinitions, HoconConvertible, 
+# RepoDefinitions, HoconConvertible,
 # expose the modules so docs can build
 # referencee： https://stackoverflow.com/questions/15115514/how-do-i-document-classes-without-the-module-name/31594545#31594545
 
@@ -30,45 +31,45 @@
 
 
 __all__ = [
- 'FeatureJoinJobParams', 
- 'FeatureGenerationJobParams', 
- 'FeathrClient', 
- 'DerivedFeature', 
- 'FeatureAnchor', 
- 'Feature', 
- 'ValueType', 
- 'WindowAggTransformation', 
- 'TypedKey', 
- 'DUMMYKEY', 
- 'BackfillTime', 
- 'MaterializationSettings', 
- 'RedisSink', 
- 'FeatureQuery', 
- 'LookupFeature', 
- 'Aggregation', 
- 'get_result_df', 
- 'AvroJsonSchema', 
- 'Source', 
- 'InputContext', 
- 'HdfsSource', 
+ 'FeatureJoinJobParams',
+ 'FeatureGenerationJobParams',
+ 'FeathrClient',
+ 'DerivedFeature',
+ 'FeatureAnchor',
+ 'Feature',
+ 'ValueType',
+ 'WindowAggTransformation',
+ 'TypedKey',
+ 'DUMMYKEY',
+ 'BackfillTime',
+ 'MaterializationSettings',
+ 'MonitoringSettings',
+ 'RedisSink',
+ 'MonitoringSqlSink',
+ 'FeatureQuery',
+ 'LookupFeature',
+ 'Aggregation',
+ 'get_result_df',
+ 'AvroJsonSchema',
+ 'Source',
+ 'InputContext',
+ 'HdfsSource',
  'KafkaConfig',
- 'KafKaSource', 
- 'ValueType', 
- 'BooleanFeatureType', 
- 'Int32FeatureType', 
- 'Int64FeatureType', 
- 'FloatFeatureType', 
- 'DoubleFeatureType', 
- 'StringFeatureType', 
+ 'KafKaSource',
+ 'ValueType',
+ 'BooleanFeatureType',
+ 'Int32FeatureType',
+ 'Int64FeatureType',
+ 'FloatFeatureType',
+ 'DoubleFeatureType',
+ 'StringFeatureType',
  'BytesFeatureType',
- 'FloatVectorFeatureType', 
- 'Int32VectorFeatureType', 
- 'Int64VectorFeatureType', 
- 'DoubleVectorFeatureType', 
- 'FeatureNameValidationError', 
- 'ObservationSettings', 
- 'FeaturePrinter', 
+ 'FloatVectorFeatureType',
+ 'Int32VectorFeatureType',
+ 'Int64VectorFeatureType',
+ 'DoubleVectorFeatureType',
+ 'FeatureNameValidationError',
+ 'ObservationSettings',
+ 'FeaturePrinter',
  'SparkExecutionConfiguration',
  ]
-
-

feathr_project/feathr/client.py

@@ -22,6 +22,7 @@
 from feathr.spark_provider.feathr_configurations import SparkExecutionConfiguration
 from feathr.definition.feature_derivations import DerivedFeature
 from feathr.definition.materialization_settings import MaterializationSettings
+from feathr.definition.monitoring_settings import MonitoringSettings
 from feathr.protobuf.featureValue_pb2 import FeatureValue
 from feathr.definition.query_feature_list import FeatureQuery
 from feathr.definition.settings import ObservationSettings
@@ -237,7 +238,7 @@ def build_features(self, anchor_list: List[FeatureAnchor] = [], derived_feature_
  self.registry.save_to_feature_config_from_context(anchor_list, derived_feature_list, self.local_workspace_dir)
  self.anchor_list = anchor_list
  self.derived_feature_list = derived_feature_list
- 
+
  # Pretty print anchor_list
  if verbose and self.anchor_list:
  FeaturePrinter.pretty_print_anchors(self.anchor_list)
@@ -400,7 +401,7 @@ def get_offline_features(self,
  observation_settings: ObservationSettings,
  feature_query: Union[FeatureQuery, List[FeatureQuery]],
  output_path: str,
- execution_configuratons: Union[SparkExecutionConfiguration ,Dict[str,str]] = None,
+ execution_configuratons: Union[SparkExecutionConfiguration ,Dict[str,str]] = {},
  udf_files = None,
  verbose: bool = False
  ):
@@ -441,15 +442,15 @@ def get_offline_features(self,
  _FeatureRegistry.save_to_feature_config_from_context(self.anchor_list, self.derived_feature_list, self.local_workspace_dir)
  else:
  raise RuntimeError("Please call FeathrClient.build_features() first in order to get offline features")
- 
+
  # Pretty print feature_query
  if verbose and feature_query:
  FeaturePrinter.pretty_print_feature_query(feature_query)
 
  write_to_file(content=config, full_file_name=config_file_path)
  return self._get_offline_features_with_config(config_file_path, execution_configuratons, udf_files=udf_files)
 
- def _get_offline_features_with_config(self, feature_join_conf_path='feature_join_conf/feature_join.conf', execution_configuratons: Dict[str,str] = None, udf_files=[]):
+ def _get_offline_features_with_config(self, feature_join_conf_path='feature_join_conf/feature_join.conf', execution_configuratons: Dict[str,str] = {}, udf_files=[]):
  """Joins the features to your offline observation dataset based on the join config.
 
  Args:
@@ -523,7 +524,16 @@ def wait_job_to_finish(self, timeout_sec: int = 300):
  else:
  raise RuntimeError('Spark job failed.')
 
- def materialize_features(self, settings: MaterializationSettings, execution_configuratons: Union[SparkExecutionConfiguration ,Dict[str,str]] = None, verbose: bool = False):
+ def monitor_features(self, settings: MonitoringSettings, execution_configuratons: Union[SparkExecutionConfiguration ,Dict[str,str]] = {}, verbose: bool = False):
+ """Create a offline job to generate statistics to monitor feature data
+
+ Args:
+ settings: Feature monitoring settings
+ execution_configuratons: a dict that will be passed to spark job when the job starts up, i.e. the "spark configurations". Note that not all of the configuration will be honored since some of the configurations are managed by the Spark platform, such as Databricks or Azure Synapse. Refer to the [spark documentation](https://spark.apache.org/docs/latest/configuration.html) for a complete list of spark configurations.
+ """
+ self.materialize_features(settings, execution_configuratons, verbose)
+
+ def materialize_features(self, settings: MaterializationSettings, execution_configuratons: Union[SparkExecutionConfiguration ,Dict[str,str]] = {}, verbose: bool = False):
  """Materialize feature data
 
  Args:
@@ -551,12 +561,12 @@ def materialize_features(self, settings: MaterializationSettings, execution_conf
  self._materialize_features_with_config(config_file_path, execution_configuratons, udf_files)
  if os.path.exists(config_file_path):
  os.remove(config_file_path)
- 
+
  # Pretty print feature_names of materialized features
  if verbose and settings:
  FeaturePrinter.pretty_print_materialize_features(settings)
 
- def _materialize_features_with_config(self, feature_gen_conf_path: str = 'feature_gen_conf/feature_gen.conf',execution_configuratons: Dict[str,str] = None, udf_files=[]):
+ def _materialize_features_with_config(self, feature_gen_conf_path: str = 'feature_gen_conf/feature_gen.conf',execution_configuratons: Dict[str,str] = {}, udf_files=[]):
  """Materializes feature data based on the feature generation config. The feature
  data will be materialized to the destination specified in the feature generation config.
 
@@ -578,24 +588,29 @@ def _materialize_features_with_config(self, feature_gen_conf_path: str = 'featur
  optional_params = []
  if _EnvVaraibleUtil.get_environment_variable('KAFKA_SASL_JAAS_CONFIG'):
  optional_params = optional_params + ['--kafka-config', self._get_kafka_config_str()]
- return self.feathr_spark_laucher.submit_feathr_job(
- job_name=self.project_name + '_feathr_feature_materialization_job',
- main_jar_path=self._FEATHR_JOB_JAR_PATH,
- python_files=cloud_udf_paths,
- main_class_name='com.linkedin.feathr.offline.job.FeatureGenJob',
- arguments=[
+ arguments = [
  '--generation-config', self.feathr_spark_laucher.upload_or_get_cloud_path(
-  generation_config.generation_config_path),
+ generation_config.generation_config_path),
  # Local Config, comma seperated file names
  '--feature-config', self.feathr_spark_laucher.upload_or_get_cloud_path(
-  generation_config.feature_config),
+ generation_config.feature_config),
  '--redis-config', self._getRedisConfigStr(),
  '--s3-config', self._get_s3_config_str(),
  '--adls-config', self._get_adls_config_str(),
  '--blob-config', self._get_blob_config_str(),
  '--sql-config', self._get_sql_config_str(),
- '--snowflake-config', self._get_snowflake_config_str()
- ] + optional_params,
+ '--snowflake-config', self._get_snowflake_config_str(),
+ ] + optional_params
+ monitoring_config_str = self._get_monitoring_config_str()
+ if monitoring_config_str:
+ arguments.append('--monitoring-config')
+ arguments.append(monitoring_config_str)
+ return self.feathr_spark_laucher.submit_feathr_job(
+ job_name=self.project_name + '_feathr_feature_materialization_job',
+ main_jar_path=self._FEATHR_JOB_JAR_PATH,
+ python_files=cloud_udf_paths,
+ main_class_name='com.linkedin.feathr.offline.job.FeatureGenJob',
+ arguments=arguments,
  reference_files_path=[],
  configuration=execution_configuratons,
  )
@@ -688,6 +703,22 @@ def _get_sql_config_str(self):
  """.format(JDBC_TABLE=table, JDBC_USER=user, JDBC_PASSWORD=password, JDBC_DRIVER = driver, JDBC_AUTH_FLAG = auth_flag, JDBC_TOKEN = token)
  return config_str
 
+ def _get_monitoring_config_str(self):
+ """Construct monitoring-related config string."""
+ url = self.envutils.get_environment_variable_with_default('monitoring', 'database', 'sql', 'url')
+ user = self.envutils.get_environment_variable_with_default('monitoring', 'database', 'sql', 'user')
+ password = self.envutils.get_environment_variable('MONITORING_DATABASE_SQL_PASSWORD')
+ if url:
+ # HOCCON format will be parsed by the Feathr job
+ config_str = """
+ MONITORING_DATABASE_SQL_URL: "{url}"
+ MONITORING_DATABASE_SQL_USER: {user}
+ MONITORING_DATABASE_SQL_PASSWORD: {password}
+ """.format(url=url, user=user, password=password)
+ return config_str
+ else:
+ ""
+
  def _get_snowflake_config_str(self):
  """Construct the Snowflake config string for jdbc. The url, user, role and other parameters can be set via
  yaml config. Password can be set via environment variables."""

feathr_project/feathr/definition/monitoring_settings.py

@@ -0,0 +1,8 @@
+from feathr.definition.materialization_settings import MaterializationSettings
+
+
+# it's completely the same as MaterializationSettings. But we renamed it to improve usability.
+# In the future, we may want to rely a separate system other than MaterializationSettings to generate stats.
+class MonitoringSettings(MaterializationSettings):
+ """Settings about monitoring features.
+ """

feathr_project/feathr/definition/sink.py

@@ -8,6 +8,27 @@ class Sink(HoconConvertible):
  """
  pass
 
+class MonitoringSqlSink(Sink):
+ """SQL-based sink that stores feature monitoring results.
+
+ Attributes:
+ table_name: output table name
+ """
+ def __init__(self, table_name: str) -> None:
+ self.table_name = table_name
+
+ def to_feature_config(self) -> str:
+ """Produce the config used in feature monitoring"""
+ tm = Template(""" 
+ {
+ name: MONITORING
+ params: {
+ table_name: "{{source.table_name}}"
+ }
+ }
+ """)
+ msg = tm.render(source=self)
+ return msg
 
 class RedisSink(Sink):
  """Redis-based sink use to store online feature data, can be used in batch job or streaming job.

feathr_project/feathr/spark_provider/_abc.py

@@ -9,7 +9,7 @@ class SparkJobLauncher(ABC):
  @abstractmethod
  def upload_or_get_cloud_path(self, local_path_or_http_path: str):
  """upload a file from local path or an http path to the current work directory. Should support transferring file from an http path to cloud working storage, or upload directly from a local storage.
- 
+
  Args:
  local_path_or_http_path (str): local path or http path
  """
@@ -19,7 +19,7 @@ def upload_or_get_cloud_path(self, local_path_or_http_path: str):
  @abstractmethod
  def submit_feathr_job(self, job_name: str, main_jar_path: str, main_class_name: str, arguments: List[str],
  reference_files_path: List[str], job_tags: Dict[str, str] = None,
- configuration: Dict[str, str] = None):
+ configuration: Dict[str, str] = {}):
  """
  Submits the feathr job
 

feathr_project/feathr/spark_provider/_databricks_submission.py

@@ -111,12 +111,12 @@ def upload_local_file(self, local_path: str) -> str:
  file_name = os.path.basename(local_path)
  # returned paths for the uploaded file
  returned_path = os.path.join(self.databricks_work_dir, file_name)
- # `local_path_or_http_path` will be either string or PathLib object, so normalize it to string 
+ # `local_path_or_http_path` will be either string or PathLib object, so normalize it to string
  local_path = str(local_path)
  DbfsApi(self.api_client).cp(recursive=True, overwrite=True, src=local_path, dst=returned_path)
  return returned_path
 
- def submit_feathr_job(self, job_name: str, main_jar_path: str, main_class_name: str, arguments: List[str], python_files: List[str], reference_files_path: List[str] = [], job_tags: Dict[str, str] = None, configuration: Dict[str, str] = None):
+ def submit_feathr_job(self, job_name: str, main_jar_path: str, main_class_name: str, arguments: List[str], python_files: List[str], reference_files_path: List[str] = [], job_tags: Dict[str, str] = None, configuration: Dict[str, str] = {}):
  """
  submit the feathr job to databricks
  Refer to the databricks doc for more details on the meaning of the parameters:
@@ -140,6 +140,10 @@ def submit_feathr_job(self, job_name: str, main_jar_path: str, main_class_name:
  submission_params['run_name'] = job_name
  if 'existing_cluster_id' not in submission_params:
  # if users don't specify existing_cluster_id
+ # Solving this issue: Handshake fails trying to connect from Azure Databricks to Azure PostgreSQL with SSL
+ # https://docs.microsoft.com/en-us/answers/questions/170730/handshake-fails-trying-to-connect-from-azure-datab.html
+ configuration['spark.executor.extraJavaOptions'] = '-Djava.security.properties='
+ configuration['spark.driver.extraJavaOptions'] = '-Djava.security.properties='
  submission_params['new_cluster']['spark_conf'] = configuration
  submission_params['new_cluster']['custom_tags'] = job_tags
  # the feathr main jar file is anyway needed regardless it's pyspark or scala spark
@@ -227,15 +231,15 @@ def get_job_tags(self) -> Dict[str, str]:
  assert self.res_job_id is not None
  # For result structure, see https://docs.microsoft.com/en-us/azure/databricks/dev-tools/api/2.0/jobs#--response-structure-6
  result = RunsApi(self.api_client).get_run(self.res_job_id)
- 
+
  if 'new_cluster' in result['cluster_spec']:
  custom_tags = result['cluster_spec']['new_cluster']['custom_tags']
  return custom_tags
  else:
  # this is not a new cluster; it's an existing cluster.
  logger.warning("Job tags are not available since you are using an existing Databricks cluster. Consider using 'new_cluster' in databricks configuration.")
  return None
- 
+
 
  def download_result(self, result_path: str, local_folder: str):
  """

feathr_project/feathr/spark_provider/_synapse_submission.py

@@ -74,7 +74,7 @@ def download_result(self, result_path: str, local_folder: str):
 
  def submit_feathr_job(self, job_name: str, main_jar_path: str = None, main_class_name: str = None, arguments: List[str] = None,
  python_files: List[str]= None, reference_files_path: List[str] = None, job_tags: Dict[str, str] = None,
- configuration: Dict[str, str] = None):
+ configuration: Dict[str, str] = {}):
  """
  Submits the feathr job
  Refer to the Apache Livy doc for more details on the meaning of the parameters:

feathr_project/test/test_azure_feature_monitoring_e2e.py

@@ -0,0 +1,25 @@
+import os
+from pathlib import Path
+
+from feathr import MonitoringSettings
+from feathr import RedisSink, MonitoringSqlSink
+
+from test_fixture import (basic_test_setup, get_online_test_table_name)
+
+
+def test_feature_monitoring():
+ online_test_table = get_online_test_table_name("nycTaxiCITable")
+ test_workspace_dir = Path(
+ __file__).parent.resolve() / "test_user_workspace"
+
+ client = basic_test_setup(os.path.join(test_workspace_dir, "feathr_config.yaml"))
+
+ monitor_sink = MonitoringSqlSink(table_name=online_test_table)
+ settings = MonitoringSettings("monitoringSetting",
+ sinks=[monitor_sink],
+ feature_names=[
+ "f_location_avg_fare", "f_location_max_fare"])
+ client.monitor_features(settings)
+ # just assume the job is successful without validating the actual result in Redis. Might need to consolidate
+ # this part with the test_feathr_online_store test case
+ client.wait_job_to_finish(timeout_sec=900)