DataDriftDetector Class

Run a backfill job over a given specified start and end date.

See https://aka.ms/datadrift for details on data drift backfill runs.

NOTE: Backfill is only supported on dataset-based DataDriftDetector objects.

backfill(start_date, end_date, compute_target=None, create_compute_target=False)

Parameters

Returns

Create a new DataDriftDetector object from a baseline tabular dataset and a target time series dataset.

static create_from_datasets(workspace, name, baseline_dataset, target_dataset, compute_target=None, frequency=None, feature_list=None, alert_config=None, drift_threshold=None, latency=None)

Parameters

Returns

Exceptions

Remarks

Dataset-based DataDriftDetectors enable you to calculate data drift between a baseline dataset, which must be a TabularDataset, and a target dataset, which must be a time series dataset. A time series dataset is simply a TabularDataset with the fine_grain_timestamp property. The DataDriftDetector can then run adhoc or scheduled jobs to determine if the target dataset has drifted from the baseline dataset.

   from azureml.core import Workspace, Dataset
   from azureml.datadrift import DataDriftDetector

   ws = Workspace.from_config()
   baseline = Dataset.get_by_name(ws, 'my_baseline_dataset')
   target = Dataset.get_by_name(ws, 'my_target_dataset')

   detector = DataDriftDetector.create_from_datasets(workspace=ws,
                                                     name="my_unique_detector_name",
                                                     baseline_dataset=baseline,
                                                     target_dataset=target,
                                                     compute_target_name='my_compute_target',
                                                     frequency="Day",
                                                     feature_list=['my_feature_1', 'my_feature_2'],
                                                     alert_config=AlertConfiguration(email_addresses=['user@contoso.com']),
                                                     drift_threshold=0.3,
                                                     latency=1)

Delete the schedule for the DataDriftDetector object.

delete(wait_for_completion=True)

Parameters

Disable the schedule for the DataDriftDetector object.

disable_schedule(wait_for_completion=True)

Parameters

Create a schedule to run dataset-based DataDriftDetector job.

enable_schedule(create_compute_target=False, wait_for_completion=True)

Parameters

Retrieve a unique DataDriftDetector object for a given workspace and name.

static get_by_name(workspace, name)

Parameters

Returns

Get a tuple of the drift results and metrics for a specific DataDriftDetector over a given time window.

get_output(start_time=None, end_time=None, run_id=None)

Parameters

Returns

Remarks

This method returns a tuple of drift results and metrics for either a time window or run ID based on the type of run: an adhoc run, a scheduled run, and a backfill run.

• To retrieve adhoc run results, there is only one way: run_id should be a valid GUID.

• To retrieve scheduled runs and backfill run results, there are two different ways: either assign a valid GUID to run_id or assign a specific start_time and/or end_time (inclusive) while keeping run_id as None.

• If run_id, start_time, and end_time are not None in the same method call, a parameter validation exception is raised.

NOTE: Specify either start_time and end_time parameters or the run_id parameter, but not both.

It's possible that there are multiple results for the same target date (target date means target dataset start date for dataset-based drift). Therefore, it's necessary to identify and handle duplicate results. For dataset-based drift, if results are for the same target date, then they are duplicated results. The get_output method will dedup any duplicated results by one rule: always pick up the latest generated results.

The get_output method can be used to retrieve all outputs or partial outputs of scheduled runs in a specific time range between start_time and end_time (boundary included). You can also limit the results of an individual adhoc by specifying the run_id.

Use the following guidelines to help interpret results returned from the get_output method:

• Principle for filtering is "overlapping": as long as there is an overlap between the actual result time (dataset-based: target dataset [start date, end date]) and the given [start_time, end_time], then the result will be picked up.

• If there are multiple outputs for one target date because the drift calculation was executed several times against that day, then only the latest output will be picked by default.

• Given there are multiple types of a data drift instance, the result contents could be various.

For dataset-based results, the output will look like:

   results : [{'drift_type': 'DatasetBased',
               'result':[{'has_drift': True, 'drift_threshold': 0.3,
                          'start_date': '2019-04-03', 'end_date': '2019-04-04',
                          'base_dataset_id': '4ac144ef-c86d-4c81-b7e5-ea6bbcd2dc7d',
                          'target_dataset_id': '13445141-aaaa-bbbb-cccc-ea23542bcaf9'}]}]
   metrics : [{'drift_type': 'DatasetBased',
               'metrics': [{'schema_version': '0.1',
                            'start_date': '2019-04-03', 'end_date': '2019-04-04',
                            'baseline_dataset_id': '4ac144ef-c86d-4c81-b7e5-ea6bbcd2dc7d',
                            'target_dataset_id': '13445141-aaaa-bbbb-cccc-ea23542bcaf9'
                            'dataset_metrics': [{'name': 'datadrift_coefficient', 'value': 0.53459}],
                            'column_metrics': [{'feature1': [{'name': 'datadrift_contribution',
                                                              'value': 288.0},
                                                             {'name': 'wasserstein_distance',
                                                              'value': 4.858040000000001},
                                                             {'name': 'energy_distance',
                                                              'value': 2.7204799576545313}]}]}]}]

Get a list of DataDriftDetector objects for the specified workspace and optional dataset.

NOTE: Passing in only the workspace parameter will return all DataDriftDetector objects, defined in the workspace.

static list(workspace, baseline_dataset=None, target_dataset=None)

Parameters

Returns

Run a single point in time data drift analysis.

run(target_date, compute_target=None, create_compute_target=False, feature_list=None, drift_threshold=None)

Parameters

Returns

Show data drift trend in given time range.

By default, this method shows the most recent 10 cycles. For example, if frequency is day, then it will be the most recent 10 days. If frequency is week, then it will be the most recent 10 weeks.

show(start_time=None, end_time=None)

Parameters

Returns

Update the schedule associated with the DataDriftDetector object.

Optional parameter values can be set to None, otherwise they default to their existing values.

update(compute_target=Ellipsis, feature_list=Ellipsis, schedule_start=Ellipsis, alert_config=Ellipsis, drift_threshold=Ellipsis, wait_for_completion=True)

Parameters

Returns

Get the alert configuration for the DataDriftDetector object.

Returns

Get the baseline dataset associated with the DataDriftDetector object.

Returns

Get the compute target attached to the DataDriftDetector object.

Returns

Get the drift threshold for the DataDriftDetector object.

Returns

Get the type of the DataDriftDetector, 'DatasetBased' is the only value supported for now.

Returns

Get the boolean value indicating whether the DataDriftDetector object is enabled.

Returns

Get the list of whitelisted features for the DataDriftDetector object.

Returns

Get the frequency of the DataDriftDetector schedule.

Returns

Get the interval of the DataDriftDetector schedule.

Returns

Get the latency of the DataDriftDetector schedule jobs (in hours).

Returns

Get the name of the DataDriftDetector object.

Returns

Get the start time of the schedule.

Returns

Denotes the state of the DataDriftDetector schedule.

Returns

Get the target dataset associated with the DataDriftDetector object.

Returns

Get the workspace of the DataDriftDetector object.

Returns