Contains an Apache Airflow provider that allows you to author Tecton workflows inside Airflow.
Two basic capabilities are supported:
- Submitting materialization jobs
- Waiting for Feature View/Feature Service data to materialize.
Note this package is in preview and it will not work with your Tecton installation unless enabled.
You can install this package via pip install airflow-tecton. Note that this requires apache-airflow>=2.0.
For a deployed environment, add airflow-tecton to your requirements.txt or wherever you configure installed packages.
You can confirm a successful installation by running airflow providers list, which should show airflow-tecton in the list.
This provider uses operators that interface with Tecton's API, thus it requires you set up Airflow to connect to Tecton.
You can add a new connection by going to Connections under the Admin tab in the Airflow UI. From there, hit the + button and select Tecton in the connection type dropdown. From there, you can configure the connection name, Tecton URL, and Tecton API key. Note that the default connection name is tecton_default, so we recommend starting with this as a connection name to minimize configuration.
A BatchFeatureView and a StreamFeatureView can be configured for manual triggering. To do so, set batch_trigger=BatchTriggerType.MANUAL. When set to manual, Tecton will not automatically create any batch materialization jobs for the Feature View.
For a StreamFeatureView, only batch materialization job scheduling will be impacted by the batch_trigger setting. Streaming materialization job scheduling will still be managed by Tecton.
Here’s an example of a BatchFeatureView configured for manual triggering.
from tecton import batch_feature_view, FilteredSource, Aggregation, BatchTriggerType
from fraud.entities import user
from fraud.data_sources.transactions import transactions_batch
from datetime import datetime, timedelta
@batch_feature_view(
sources=[FilteredSource(transactions_batch)],
entities=[user],
mode='spark_sql',
aggregation_interval=timedelta(days=1),
aggregations=[
Aggregation(column='transaction', function='count', time_window=timedelta(days=1)),
Aggregation(column='transaction', function='count', time_window=timedelta(days=30)),
Aggregation(column='transaction', function='count', time_window=timedelta(days=90))
],
online=False,
offline=True,
feature_start_time=datetime(2022, 5, 1),
tags={'release': 'production'},
owner='matt@tecton.ai',
description='User transaction totals over a series of time windows, updated daily.',
batch_trigger=BatchTriggerType.MANUAL # Use manual triggers
)
def user_transaction_counts(transactions):
return f'''
SELECT
user_id,
1 as transaction,
timestamp
FROM
{transactions}
'''If a Data Source input to the Feature View has data_delay set, then that delay will still be factored in to constructing training data sets but does not impact when the job can be triggered with the materialization API.
There are two methods available to submit materialization jobs:
- TectonTriggerOperator: This triggers a materialization job for a Feature View. Tecton will retry any failing jobs automatically. Note that completion of this operator only means submission succeeded. To wait for completion, you must combine this with
TectonSensor. - TectonJobOperator: This triggers a materialization job for a Feature View with no retries. Additionally, when this operator is terminated, it will make a best effort to clean up the execution of the materialization job. Using this operator allows you to use standard Airflow keyword arguments to configure retry behavior. Additionally, this operator is synchronous, meaning that when the operator has succeeded, the underlying job has succeeded.
Both of these require the following arguments:
- workspace - the workspace name of the Feature View you intend to materialize
- feature_view - the name of the Feature View you intend to materialize
- online - whether the job should materialize to the online store. This requires that your FeatureView also has online materialization enabled.
- offline - whether the job should materialize to the offline store. This requires that your FeatureView also has offline materialization enabled.
The time interval of the materialization job is configured automatically using Airflow templates. By default, it is from the data_interval_start to the data_interval_end of your DAG run. These can overridden if necessary.
from airflow_tecton import TectonJobOperator, TectonTriggerOperator
TectonJobOperator(
task_id="tecton_job",
workspace="my_workspace",
feature_view="my_fv",
online=False,
offline=True,
retries=3
)
TectonTriggerOperator(
task_id="trigger_tecton",
workspace="my_workspace",
feature_view="my_fv",
online=True,
offline=True,
)This enables you to wait for Materialization to complete for both Feature Views and Feature Services. Common uses are for monitoring as well as kicking off a training job after daily materialization completes.
from airflow_tecton import TectonSensor
TectonSensor(
task_id="wait_for_fs_online",
workspace="my_workspace",
feature_service="my_fs",
online=True,
offline=False,
)
TectonSensor(
task_id="wait_for_fv",
workspace="my_workspace",
feature_view="my_fv",
online=True,
offline=True,
)See example dags here.
This repo uses pre-commit. Run pre-commit install in the repo root to configure pre-commit hooks. Pre-commit hooks take care of running unit tests as well as linting files.
Run python -m pytest tests/ in the repo root.
This is licensed with the Apache 2.0 License.