def monitor_frequency(interval=frequency, hour_interval=None, starting_hour=None):
# this allows users to define the frequency of data drift monitoring
if interval == "daily":
monitoring_frequency = CronExpressionGenerator.daily()
if interval == "hourly":
monitoring_frequency = CronExpressionGenerator.hourly()
if interval == "others":
monitoring_frequency = CronExpressionGenerator.daily_every_x_hours(
hour_interval, starting_hour
)
return monitoring_frequency
from tests.integ.retry import retries
XGBOOST_DATA_PATH = os.path.join(DATA_DIR, "xgboost_model")
ENDPOINT_INPUT_LOCAL_PATH = "/opt/ml/processing/input/endpoint"
PROBABILITY_THRESHOLD = 0.5005
PROBLEM_TYPE = "Regression"
INFERENCE_ATTRIBUTE = "0"
HEADER_OF_LABEL = "Label"
HEADER_OF_PREDICTED_LABEL = "Prediction"
HEADERS_OF_FEATURES = ["F1", "F2", "F3", "F4", "F5", "F6", "F7"]
ALL_HEADERS = [
*HEADERS_OF_FEATURES, HEADER_OF_LABEL, HEADER_OF_PREDICTED_LABEL
]
CRON = "cron(0 * * * ? *)"
UPDATED_CRON = CronExpressionGenerator.daily()
MAX_RUNTIME_IN_SECONDS = 30 * 60
UPDATED_MAX_RUNTIME_IN_SECONDS = 25 * 60
ROLE = "SageMakerRole"
INSTANCE_COUNT = 1
INSTANCE_TYPE = "ml.c5.xlarge"
VOLUME_SIZE_IN_GB = 100
START_TIME_OFFSET = "-PT1H"
END_TIME_OFFSET = "-PT0H"
TEST_TAGS = [{"Key": "integration", "Value": "test"}]
# TODO: Remove this workaround once the API service fix is deployed to Prod
TEST_ENV = {"problem_type": PROBLEM_TYPE}
@pytest.yield_fixture(scope="module")
def endpoint_name(sagemaker_session):
def main(resources, train_data):
# configurarion
AWS_DEFAULT_REGION = os.getenv('AWS_DEFAULT_REGION', 'eu-west-1')
AWS_PROFILE = os.getenv('AWS_PROFILE', 'default')
AWS_ACCESS_KEY_ID = os.getenv('AWS_ACCESS_KEY_ID', None)
AWS_SECRET_ACCESS_KEY = os.getenv('AWS_SECRET_ACCESS_KEY', None)
b3_session, sm_client, sm_runtime, sm_session = get_sm_session(
region=AWS_DEFAULT_REGION,
profile_name=AWS_PROFILE,
aws_access_key_id=AWS_ACCESS_KEY_ID,
aws_secret_access_key=AWS_SECRET_ACCESS_KEY)
BASE_JOB_PREFIX = os.getenv('BASE_JOB_PREFIX', 'sts')
ROLE_ARN = os.getenv('AWS_ROLE', sagemaker.get_execution_role())
outputs = resources
bucket = sm_session.default_bucket()
prefix = "{}/{}".format(BASE_JOB_PREFIX, resources['endpoint']['name'])
if 'monitor' not in resources:
raise ValueError("Monitoring not enabled")
if 's3_capture_upload_path' not in resources['monitor']:
raise ValueError("Monitoring not enabled")
baseline_prefix = prefix + "/baselining"
baseline_data_prefix = baseline_prefix + "/data"
baseline_results_prefix = baseline_prefix + "/results"
baseline_data_uri = "s3://{}/{}".format(bucket, baseline_data_prefix)
baseline_results_uri = "s3://{}/{}".format(bucket, baseline_results_prefix)
outputs['monitor'].update({
'baseline': {
'data_uri': baseline_data_uri,
'results_uri': baseline_results_uri
}
})
_l.info("Baseline data uri: {}".format(baseline_data_uri))
_l.info("Baseline results uri: {}".format(baseline_results_uri))
ground_truth_upload_path = f"s3://{bucket}/{prefix}/ground_truth_data"
_l.info(f"Ground truth uri: {ground_truth_upload_path}")
outputs['monitor'].update({'ground truth uri': ground_truth_upload_path})
# Create a baselining job with training dataset
_l.info("Executing a baselining job with training dataset")
_l.info(f"baseline_data_uri: {train_data['baseline']['validate']}")
my_monitor = ModelQualityMonitor(
role=ROLE_ARN,
sagemaker_session=sm_session,
max_runtime_in_seconds=1800 # 30 minutes
)
my_monitor.suggest_baseline(
baseline_dataset=train_data['baseline']['validate'] + "/baseline.csv",
dataset_format=DatasetFormat.csv(header=True),
problem_type="Regression",
inference_attribute="prediction",
ground_truth_attribute="label",
output_s3_uri=baseline_results_uri,
wait=True)
baseline_job = my_monitor.latest_baselining_job
_l.info("suggested baseline contrains")
_l.info(
pprint.pformat(baseline_job.suggested_constraints().
body_dict["regression_constraints"]))
_l.info("suggested baseline statistics")
_l.info(
pprint.pformat(baseline_job.baseline_statistics().
body_dict["regression_metrics"]))
monitor_schedule_name = (
f"{BASE_JOB_PREFIX}-mq-sch-{datetime.datetime.utcnow():%Y-%m-%d-%H%M}")
_l.info(f"Monitoring schedule name: {monitor_schedule_name}")
outputs['monitor'].update({'schedule_name': monitor_schedule_name})
endpointInput = EndpointInput(
resources['endpoint']['name'],
"/opt/ml/processing/input_data",
inference_attribute='0' # REVIEW:
)
my_monitor.create_monitoring_schedule(
monitor_schedule_name=monitor_schedule_name,
endpoint_input=endpointInput,
output_s3_uri=baseline_results_uri,
problem_type="Regression",
ground_truth_input=ground_truth_upload_path,
constraints=baseline_job.suggested_constraints(),
# run the scheduler hourly
schedule_cron_expression=CronExpressionGenerator.hourly(),
enable_cloudwatch_metrics=True,
)
mq_schedule_details = my_monitor.describe_schedule()
while mq_schedule_details['MonitoringScheduleStatus'] == 'Pending':
_l.info(f'Waiting for {monitor_schedule_name}')
time.sleep(3)
mq_schedule_details = my_monitor.describe_schedule()
_l.debug(
f"Model Quality Monitor - schedule details: {pprint.pformat(mq_schedule_details)}"
)
_l.info(
f"Model Quality Monitor - schedule status: {mq_schedule_details['MonitoringScheduleStatus']}"
)
# save outputs to a file
with open('deploymodel_out.json', 'w') as f:
json.dump(outputs, f, default=json_default)
print('Model data baseline suggested at {}'.format(baseline_results_uri))
import datetime as datetime
from time import gmtime, strftime
mon_schedule_name = '{}-{}'.format(mon_schedule_name_base, datetime.datetime.now().strftime("%Y-%m-%d-%H%M%S"))
s3_report_path = f's3://{bucket}/{prefix}/monitor/report'
# Setup daily Cron job schedule
print(f"Attempting to create monitoring schedule as {mon_schedule_name} \n")
try:
my_default_monitor.create_monitoring_schedule(
monitor_schedule_name=mon_schedule_name,
endpoint_input=endpoint_name,
output_s3_uri=s3_report_path,
statistics=my_default_monitor.baseline_statistics(),
constraints=my_default_monitor.suggested_constraints(),
schedule_cron_expression=CronExpressionGenerator.daily(),
enable_cloudwatch_metrics=True,
)
desc_schedule_result = my_default_monitor.describe_schedule()
print('Created monitoring schedule. Schedule status: {}'.format(desc_schedule_result['MonitoringScheduleStatus']))
except:
my_default_monitor.update_monitoring_schedule(
endpoint_input=endpoint_name,
schedule_cron_expression=CronExpressionGenerator.daily()
)
print("Monitoring schedule already exists for endpoint. Updating schedule.")
def test_cron_expression_generator_daily_every_x_hours_returns_expected_value_when_called_with_customizations():
assert (
CronExpressionGenerator.daily_every_x_hours(hour_interval=7, starting_hour=8)
== "cron(0 8/7 ? * * *)"
)
def test_cron_expression_generator_daily_every_x_hours_returns_expected_value_when_called_without_customizations():
assert CronExpressionGenerator.daily_every_x_hours(hour_interval=6) == "cron(0 0/6 ? * * *)"
def test_cron_expression_generator_daily_returns_expected_value_when_called_with_parameters():
assert CronExpressionGenerator.daily(hour=5) == "cron(0 5 ? * * *)"
def test_cron_expression_generator_hourly_returns_expected_value():
assert CronExpressionGenerator.hourly() == "cron(0 * ? * * *)"
max_runtime_in_seconds=3600,
)
def create_baseline():
print(f'Baseline data uri: {baseline_data_uri}')
print(f'Baseline results uri: {baseline_results_uri}')
my_default_monitor.suggest_baseline(
baseline_dataset=baseline_data_uri,
dataset_format=DatasetFormat.csv(header=False),
output_s3_uri=baseline_results_uri,
wait=True
)
mon_schedule_name = 'xgb-boston-pred-model-monitor-schedule-' + strftime("%Y-%m-%d-%H-%M-%S", gmtime())
my_default_monitor.create_monitoring_schedule(
monitor_schedule_name=mon_schedule_name,
endpoint_input=endpoint_name,
output_s3_uri=baseline_results_uri.replace('baseline_results', 'monitor_reports'),
statistics=baseline_results_uri + '/statistics.json',
constraints=baseline_results_uri + '/constraints.json',
schedule_cron_expression=CronExpressionGenerator.hourly(),
enable_cloudwatch_metrics=True,
)