class TestDataflowPythonOperator(unittest.TestCase):
def setUp(self):
self.dataflow = DataflowCreatePythonJobOperator(
task_id=TASK_ID,
py_file=PY_FILE,
job_name=JOB_NAME,
py_options=PY_OPTIONS,
dataflow_default_options=DEFAULT_OPTIONS_PYTHON,
options=ADDITIONAL_OPTIONS,
poll_sleep=POLL_SLEEP,
location=TEST_LOCATION,
)
def test_init(self):
"""Test DataFlowPythonOperator instance is properly initialized."""
self.assertEqual(self.dataflow.task_id, TASK_ID)
self.assertEqual(self.dataflow.job_name, JOB_NAME)
self.assertEqual(self.dataflow.py_file, PY_FILE)
self.assertEqual(self.dataflow.py_options, PY_OPTIONS)
self.assertEqual(self.dataflow.py_interpreter, PY_INTERPRETER)
self.assertEqual(self.dataflow.poll_sleep, POLL_SLEEP)
self.assertEqual(self.dataflow.dataflow_default_options,
DEFAULT_OPTIONS_PYTHON)
self.assertEqual(self.dataflow.options, EXPECTED_ADDITIONAL_OPTIONS)
@mock.patch(
'airflow.providers.google.cloud.operators.dataflow.DataflowHook')
@mock.patch('airflow.providers.google.cloud.operators.dataflow.GCSHook')
def test_exec(self, gcs_hook, dataflow_mock):
"""Test DataflowHook is created and the right args are passed to
start_python_workflow.
"""
start_python_hook = dataflow_mock.return_value.start_python_dataflow
gcs_provide_file = gcs_hook.return_value.provide_file
self.dataflow.execute(None)
self.assertTrue(dataflow_mock.called)
expected_options = {
'project': 'test',
'staging_location': 'gs://test/staging',
'output': 'gs://test/output',
'labels': {
'foo': 'bar',
'airflow-version': TEST_VERSION
},
}
gcs_provide_file.assert_called_once_with(object_url=PY_FILE)
start_python_hook.assert_called_once_with(
job_name=JOB_NAME,
variables=expected_options,
dataflow=mock.ANY,
py_options=PY_OPTIONS,
py_interpreter=PY_INTERPRETER,
py_requirements=None,
py_system_site_packages=False,
on_new_job_id_callback=mock.ANY,
project_id=None,
location=TEST_LOCATION,
)
self.assertTrue(self.dataflow.py_file.startswith('/tmp/dataflow'))
def setUp(self):
self.dataflow = DataflowCreatePythonJobOperator(
task_id=TASK_ID,
py_file=PY_FILE,
job_name=JOB_NAME,
py_options=PY_OPTIONS,
dataflow_default_options=DEFAULT_OPTIONS_PYTHON,
options=ADDITIONAL_OPTIONS,
poll_sleep=POLL_SLEEP)
def setUp(self):
self.dataflow = DataflowCreatePythonJobOperator(
task_id=TASK_ID,
py_file=PY_FILE,
job_name=JOB_NAME,
py_options=PY_OPTIONS,
dataflow_default_options=DEFAULT_OPTIONS_PYTHON,
options=ADDITIONAL_OPTIONS,
poll_sleep=POLL_SLEEP,
location=TEST_LOCATION,
)
self.expected_airflow_version = 'v' + airflow.version.version.replace(
".", "-").replace("+", "-")
jar_to_local >> start_java_job_local
with models.DAG(
"example_gcp_dataflow_native_python",
default_args=default_args,
schedule_interval=None, # Override to match your needs
tags=['example'],
) as dag_native_python:
# [START howto_operator_start_python_job]
start_python_job = DataflowCreatePythonJobOperator(
task_id="start-python-job",
py_file=GCS_PYTHON,
py_options=[],
job_name='{{task.task_id}}',
options={
'output': GCS_OUTPUT,
},
py_requirements=['apache-beam[gcp]==2.21.0'],
py_interpreter='python3',
py_system_site_packages=False,
location='europe-west3')
# [END howto_operator_start_python_job]
start_python_job_local = DataflowCreatePythonJobOperator(
task_id="start-python-job-local",
py_file='apache_beam.examples.wordcount',
py_options=['-m'],
job_name='{{task.task_id}}',
options={
'output': GCS_OUTPUT,
},
def create_evaluate_ops( # pylint: disable=too-many-arguments
task_prefix: str,
data_format: str,
input_paths: List[str],
prediction_path: str,
metric_fn_and_keys: Tuple[T, Iterable[str]],
validate_fn: T,
batch_prediction_job_id: Optional[str] = None,
region: Optional[str] = None,
project_id: Optional[str] = None,
dataflow_options: Optional[Dict] = None,
model_uri: Optional[str] = None,
model_name: Optional[str] = None,
version_name: Optional[str] = None,
dag: Optional[DAG] = None,
py_interpreter="python3",
):
"""
Creates Operators needed for model evaluation and returns.
It gets prediction over inputs via Cloud ML Engine BatchPrediction API by
calling MLEngineBatchPredictionOperator, then summarize and validate
the result via Cloud Dataflow using DataFlowPythonOperator.
For details and pricing about Batch prediction, please refer to the website
https://cloud.google.com/ml-engine/docs/how-tos/batch-predict
and for Cloud Dataflow, https://cloud.google.com/dataflow/docs/
It returns three chained operators for prediction, summary, and validation,
named as <prefix>-prediction, <prefix>-summary, and <prefix>-validation,
respectively.
(<prefix> should contain only alphanumeric characters or hyphen.)
The upstream and downstream can be set accordingly like:
pred, _, val = create_evaluate_ops(...)
pred.set_upstream(upstream_op)
...
downstream_op.set_upstream(val)
Callers will provide two python callables, metric_fn and validate_fn, in
order to customize the evaluation behavior as they wish.
- metric_fn receives a dictionary per instance derived from json in the
batch prediction result. The keys might vary depending on the model.
It should return a tuple of metrics.
- validation_fn receives a dictionary of the averaged metrics that metric_fn
generated over all instances.
The key/value of the dictionary matches to what's given by
metric_fn_and_keys arg.
The dictionary contains an additional metric, 'count' to represent the
total number of instances received for evaluation.
The function would raise an exception to mark the task as failed, in a
case the validation result is not okay to proceed (i.e. to set the trained
version as default).
Typical examples are like this:
.. code-block:: python
def get_metric_fn_and_keys():
import math # imports should be outside of the metric_fn below.
def error_and_squared_error(inst):
label = float(inst['input_label'])
classes = float(inst['classes']) # 0 or 1
err = abs(classes-label)
squared_err = math.pow(classes-label, 2)
return (err, squared_err) # returns a tuple.
return error_and_squared_error, ['err', 'mse'] # key order must match.
def validate_err_and_count(summary):
if summary['err'] > 0.2:
raise ValueError('Too high err>0.2; summary=%s' % summary)
if summary['mse'] > 0.05:
raise ValueError('Too high mse>0.05; summary=%s' % summary)
if summary['count'] < 1000:
raise ValueError('Too few instances<1000; summary=%s' % summary)
return summary
For the details on the other BatchPrediction-related arguments (project_id,
job_id, region, data_format, input_paths, prediction_path, model_uri),
please refer to MLEngineBatchPredictionOperator too.
:param task_prefix: a prefix for the tasks. Only alphanumeric characters and
hyphen are allowed (no underscores), since this will be used as dataflow
job name, which doesn't allow other characters.
:type task_prefix: str
:param data_format: either of 'TEXT', 'TF_RECORD', 'TF_RECORD_GZIP'
:type data_format: str
:param input_paths: a list of input paths to be sent to BatchPrediction.
:type input_paths: list[str]
:param prediction_path: GCS path to put the prediction results in.
:type prediction_path: str
:param metric_fn_and_keys: a tuple of metric_fn and metric_keys:
- metric_fn is a function that accepts a dictionary (for an instance),
and returns a tuple of metric(s) that it calculates.
- metric_keys is a list of strings to denote the key of each metric.
:type metric_fn_and_keys: tuple of a function and a list[str]
:param validate_fn: a function to validate whether the averaged metric(s) is
good enough to push the model.
:type validate_fn: function
:param batch_prediction_job_id: the id to use for the Cloud ML Batch
prediction job. Passed directly to the MLEngineBatchPredictionOperator as
the job_id argument.
:type batch_prediction_job_id: str
:param project_id: the Google Cloud project id in which to execute
Cloud ML Batch Prediction and Dataflow jobs. If None, then the `dag`'s
`default_args['project_id']` will be used.
:type project_id: str
:param region: the Google Cloud region in which to execute Cloud ML
Batch Prediction and Dataflow jobs. If None, then the `dag`'s
`default_args['region']` will be used.
:type region: str
:param dataflow_options: options to run Dataflow jobs. If None, then the
`dag`'s `default_args['dataflow_default_options']` will be used.
:type dataflow_options: dictionary
:param model_uri: GCS path of the model exported by Tensorflow using
tensorflow.estimator.export_savedmodel(). It cannot be used with
model_name or version_name below. See MLEngineBatchPredictionOperator for
more detail.
:type model_uri: str
:param model_name: Used to indicate a model to use for prediction. Can be
used in combination with version_name, but cannot be used together with
model_uri. See MLEngineBatchPredictionOperator for more detail. If None,
then the `dag`'s `default_args['model_name']` will be used.
:type model_name: str
:param version_name: Used to indicate a model version to use for prediction,
in combination with model_name. Cannot be used together with model_uri.
See MLEngineBatchPredictionOperator for more detail. If None, then the
`dag`'s `default_args['version_name']` will be used.
:type version_name: str
:param dag: The `DAG` to use for all Operators.
:type dag: airflow.models.DAG
:param py_interpreter: Python version of the beam pipeline.
If None, this defaults to the python3.
To track python versions supported by beam and related
issues check: https://issues.apache.org/jira/browse/BEAM-1251
:type py_interpreter: str
:returns: a tuple of three operators, (prediction, summary, validation)
:rtype: tuple(DataFlowPythonOperator, DataFlowPythonOperator,
PythonOperator)
"""
batch_prediction_job_id = batch_prediction_job_id or ""
dataflow_options = dataflow_options or {}
region = region or ""
# Verify that task_prefix doesn't have any special characters except hyphen
# '-', which is the only allowed non-alphanumeric character by Dataflow.
if not re.match(r"^[a-zA-Z][-A-Za-z0-9]*$", task_prefix):
raise AirflowException(
"Malformed task_id for DataFlowPythonOperator (only alphanumeric "
"and hyphens are allowed but got: " + task_prefix
)
metric_fn, metric_keys = metric_fn_and_keys
if not callable(metric_fn):
raise AirflowException("`metric_fn` param must be callable.")
if not callable(validate_fn):
raise AirflowException("`validate_fn` param must be callable.")
if dag is not None and dag.default_args is not None:
default_args = dag.default_args
project_id = project_id or default_args.get('project_id')
region = region or default_args['region']
model_name = model_name or default_args.get('model_name')
version_name = version_name or default_args.get('version_name')
dataflow_options = dataflow_options or default_args.get('dataflow_default_options')
evaluate_prediction = MLEngineStartBatchPredictionJobOperator(
task_id=(task_prefix + "-prediction"),
project_id=project_id,
job_id=batch_prediction_job_id,
region=region,
data_format=data_format,
input_paths=input_paths,
output_path=prediction_path,
uri=model_uri,
model_name=model_name,
version_name=version_name,
dag=dag,
)
metric_fn_encoded = base64.b64encode(dill.dumps(metric_fn, recurse=True)).decode()
evaluate_summary = DataflowCreatePythonJobOperator(
task_id=(task_prefix + "-summary"),
py_file=os.path.join(os.path.dirname(__file__), 'mlengine_prediction_summary.py'),
dataflow_default_options=dataflow_options,
options={
"prediction_path": prediction_path,
"metric_fn_encoded": metric_fn_encoded,
"metric_keys": ','.join(metric_keys),
},
py_interpreter=py_interpreter,
py_requirements=['apache-beam[gcp]>=2.14.0'],
dag=dag,
)
evaluate_summary.set_upstream(evaluate_prediction)
def apply_validate_fn(*args, templates_dict, **kwargs):
prediction_path = templates_dict["prediction_path"]
scheme, bucket, obj, _, _ = urlsplit(prediction_path)
if scheme != "gs" or not bucket or not obj:
raise ValueError("Wrong format prediction_path: {}".format(prediction_path))
summary = os.path.join(obj.strip("/"), "prediction.summary.json")
gcs_hook = GCSHook()
summary = json.loads(gcs_hook.download(bucket, summary))
return validate_fn(summary)
evaluate_validation = PythonOperator(
task_id=(task_prefix + "-validation"),
python_callable=apply_validate_fn,
templates_dict={"prediction_path": prediction_path},
dag=dag,
)
evaluate_validation.set_upstream(evaluate_summary)
return evaluate_prediction, evaluate_summary, evaluate_validation
class TestDataflowPythonOperator(unittest.TestCase):
def setUp(self):
self.dataflow = DataflowCreatePythonJobOperator(
task_id=TASK_ID,
py_file=PY_FILE,
job_name=JOB_NAME,
py_options=PY_OPTIONS,
dataflow_default_options=DEFAULT_OPTIONS_PYTHON,
options=ADDITIONAL_OPTIONS,
poll_sleep=POLL_SLEEP,
location=TEST_LOCATION,
)
self.expected_airflow_version = 'v' + airflow.version.version.replace(
".", "-").replace("+", "-")
def test_init(self):
"""Test DataFlowPythonOperator instance is properly initialized."""
assert self.dataflow.task_id == TASK_ID
assert self.dataflow.job_name == JOB_NAME
assert self.dataflow.py_file == PY_FILE
assert self.dataflow.py_options == PY_OPTIONS
assert self.dataflow.py_interpreter == PY_INTERPRETER
assert self.dataflow.poll_sleep == POLL_SLEEP
assert self.dataflow.dataflow_default_options == DEFAULT_OPTIONS_PYTHON
assert self.dataflow.options == EXPECTED_ADDITIONAL_OPTIONS
@mock.patch(
'airflow.providers.google.cloud.operators.dataflow.process_line_and_extract_dataflow_job_id_callback'
)
@mock.patch('airflow.providers.google.cloud.operators.dataflow.BeamHook')
@mock.patch(
'airflow.providers.google.cloud.operators.dataflow.DataflowHook')
@mock.patch('airflow.providers.google.cloud.operators.dataflow.GCSHook')
def test_exec(self, gcs_hook, dataflow_hook_mock, beam_hook_mock,
mock_callback_on_job_id):
"""Test DataflowHook is created and the right args are passed to
start_python_workflow.
"""
start_python_mock = beam_hook_mock.return_value.start_python_pipeline
gcs_provide_file = gcs_hook.return_value.provide_file
job_name = dataflow_hook_mock.return_value.build_dataflow_job_name.return_value
self.dataflow.execute(None)
beam_hook_mock.assert_called_once_with(runner="DataflowRunner")
self.assertTrue(self.dataflow.py_file.startswith('/tmp/dataflow'))
gcs_provide_file.assert_called_once_with(object_url=PY_FILE)
mock_callback_on_job_id.assert_called_once_with(
on_new_job_id_callback=mock.ANY)
dataflow_hook_mock.assert_called_once_with(
gcp_conn_id="google_cloud_default",
delegate_to=mock.ANY,
poll_sleep=POLL_SLEEP,
impersonation_chain=None,
drain_pipeline=False,
cancel_timeout=mock.ANY,
wait_until_finished=None,
)
expected_options = {
"project": dataflow_hook_mock.return_value.project_id,
"staging_location": 'gs://test/staging',
"job_name": job_name,
"region": TEST_LOCATION,
'output': 'gs://test/output',
'labels': {
'foo': 'bar',
'airflow-version': self.expected_airflow_version
},
}
start_python_mock.assert_called_once_with(
variables=expected_options,
py_file=gcs_provide_file.return_value.__enter__.return_value.name,
py_options=PY_OPTIONS,
py_interpreter=PY_INTERPRETER,
py_requirements=None,
py_system_site_packages=False,
process_line_callback=mock_callback_on_job_id.return_value,
)
dataflow_hook_mock.return_value.wait_for_done.assert_called_once_with(
job_id=mock.ANY,
job_name=job_name,
location=TEST_LOCATION,
multiple_jobs=False,
)
assert self.dataflow.py_file.startswith('/tmp/dataflow')
