What is an on-demand transformation?
An on-demand transformation is a feature function that is used to compute an on-demand feature. It is run both in an online inference pipeline and in a feature pipeline. In the online inference pipeline, the on-demand feature function takes as input data only available at request-time (it may also use pre-computed features from the feature store), while in the feature pipeline the on-demand feature function takes as input historical data.
Examples of on-demand transformations:
- Calculate the time since the last user interaction.
- Compute the current temperature based on a live sensor data reading.
- Derive the current zipcode of a user from their GPS coordinates.
Here is a short snippet showcasing an on-demand transformation that calculates the time since the last user interaction in an online inference pipeline:
import time
def time_since_last_interaction(last_interaction_timestamp):
current_time = time.time()
time_diff = current_time - last_interaction_timestamp
return time_diff
precomputed_features=feature_view.get_feature_vector({"id": 123})
time_since_last = time_since_last_interaction(
precomputed_features[0]) # 'last_interaction' at offset '0'
print("Time since last interaction (seconds):", time_since_last)
This example demonstrates a simple on-demand transformation (feature function) that computes the time since the last user interaction based on the current time and the timestamp of the last interaction (a lag feature).
This feature function can also be applied both in the feature pipeline using historical data, as shown below:
# import the same code for the on-demand feature fn as
# used in the online inference pipeline
import time_since_last_interaction
# read Pandas DataFrame from historical store
user_logins_df = pd.read_csv("user_logins.csv")
# Use the on-demand feature function to compute the feature
user_logins_df['time_since_last'] =
user_logins_df.apply(lambda x: time_since_last_interaction(
user_logins_df['last_login']), axis=1)
.svg)