Compute features on demand using Python user-defined functions

This article describes how to create and use on-demand features in Databricks.

Machine learning models for real-time applications often require the most recent feature values. In the example shown in the diagram, one feature for a restaurant recommendation model is the user’s current distance from a restaurant. This feature must be calculated “on demand”—that is, at the time of the scoring request. Upon receiving a scoring request, the model looks up the restaurant’s location, and then applies a pre-defined function to calculate the distance between the user’s current location and the restaurant. That distance is passed as an input to the model, along with other precomputed features from the feature store.

compute features on demand workflow

To use on-demand features, your workspace must be enabled for Unity Catalog and you must use Databricks Runtime 13.3 LTS ML or above.

What are on-demand features?

“On-demand” refers to features whose values are not known ahead of time, but are calculated at the time of inference. In Databricks, you use Python user-defined functions (UDFs) to specify how to calculate on-demand features. These functions are governed by Unity Catalog and discoverable through Catalog Explorer.


To compute features on-demand, you specify a Python user-defined function (UDF) that describes how to calculate the feature values.

  • During training, you provide this function and its input bindings in the feature_lookups parameter of the create_training_set API.

  • You must log the trained model using the Feature Store method log_model. This ensures that the model automatically evaluates on-demand features when it is used for inference.

  • For batch scoring, the score_batch API automatically calculates and returns all feature values, including on-demand features.

  • When you serve a model with Databricks Model Serving, the model automatically uses the Python UDF to compute on-demand features for each scoring request.

Create a Python UDF

You can create a Python UDF in a notebook or in Databricks SQL.

For example, running the following code in a notebook cell creates the Python UDF example_feature in the catalog main and schema default.

CREATE FUNCTION main.default.example_feature(x INT, y INT)
COMMENT 'add two numbers'
AS $$
def add_numbers(n1: int, n2: int) -> int:
  return n1 + n2

return add_numbers(x, y)

After running the code, you can navigate through the three-level namespace in Catalog Explorer to view the function definition:

function in Catalog Explorer

For more details about creating Python UDFs, see Register a Python UDF to Unity Catalog and the SQL language manual.

How to handle missing feature values

When a Python UDF depends on the result of a FeatureLookup, the value returned if the requested lookup key is not found depends on the environment. When using score_batch, the value returned is None. When using online serving, the value returned is float("nan").

The following code is an example of how to handle both cases.

import numpy as np
if x is None or np.isnan(x):
  return 0
return x * x

Train a model using on-demand features

To train the model, you use a FeatureFunction, which is passed to the create_training_set API in the feature_lookups parameter.

The following example code uses the Python UDF main.default.example_feature that was defined in the previous section.

# Install databricks-feature-engineering first with:
# %pip install databricks-feature-engineering
# dbutils.library.restartPython()

from databricks.feature_engineering import FeatureEngineeringClient
from databricks.feature_engineering import FeatureFunction, FeatureLookup
from sklearn import linear_model

fe = FeatureEngineeringClient()

features = [
  # The feature 'on_demand_feature' is computed as the sum of the the input value 'new_source_input'
  # and the pre-materialized feature 'materialized_feature_value'.
  # - 'new_source_input' must be included in base_df and also provided at inference time.
  #   - For batch inference, it must be included in the DataFrame passed to 'FeatureEngineeringClient.score_batch'.
  #   - For real-time inference, it must be included in the request.
  # - 'materialized_feature_value' is looked up from a feature table.

      udf_name="main.default.example_feature",    # UDF must be in Unity Catalog so uses a three-level namespace
        "x": "new_source_input",
        "y": "materialized_feature_value"
  # retrieve the prematerialized feature
    table_name = 'main.default.table',
    feature_names = ['materialized_feature_value'],
    lookup_key = 'id'

# base_df includes the columns 'id', 'new_source_input', and 'label'
training_set = fe.create_training_set(
  exclude_columns=['id', 'new_source_input', 'materialized_feature_value']     # drop the columns not used for training

# The training set contains the columns 'on_demand_feature' and 'label'.
training_df = training_set.load_df().toPandas()

# training_df columns ['materialized_feature_value', 'label']
X_train = training_df.drop(['label'], axis=1)
y_train = training_df.label

model = linear_model.LinearRegression().fit(X_train, y_train)

Log the model and register it to Unity Catalog

Models packaged with feature metadata can be registered to Unity Catalog. The feature tables used to create the model must be stored in Unity Catalog.

To ensure that the model automatically evaluates on-demand features when it is used for inference, you must set the registry URI and then log the model, as follows:

import mlflow


If the Python UDF that defines the on-demand features imports any Python packages, you must specify these packages using the argument extra_pip_requirements. For example:

import mlflow



On-demand features can output all data types supported by Feature Store except MapType and ArrayType.

Notebook examples: On-demand features

The following notebook shows an example of how to train and score a model that uses an on-demand feature.

Basic on-demand features demo notebook

Open notebook in new tab

The following notebook shows an example of a restaurant recommendation model. The restaurant’s location is looked up from a pre-materialized feature table published to an online store. The user’s current location is sent as part of the scoring request. The model uses an on-demand feature to compute the real-time distance from the user to the restaurant. That distance is then used as an input to the model.

Restaurant recommendation on-demand features demo notebook

Open notebook in new tab