​
Serving for Inference

When a feature is defined, it is materialized into the inference store associated with the definition. If a feature has a timestamp, only its most recent value is stored in the inference store. It can be thought of as a feature cache for inference time.

To serve a feature we initialize a serving client, specify the names and variants of the features we wish to serve, and provide the entity whos feature values to look up. Underneath the hood, Featureform will keep a mapping between a feature’s name and variant and the tables that they are actually stored in. This allows user’s to stick to the Featureform level of abstraction.

fpf = client.features([("fpf", "quickstart")], {"passenger": "1"})

​
On-Demand Features

On-demand features do not require any materialization, they are calculated at serving time. You can define your on-demand features by simply adding the ondemand_feature decorator to your function. The function will need to have serving_client, params, entities, as part of it’s definition.

import featureform as ff

@ff.ondemand_feature(name="gender_category", variant="quickstart")
def gender_category(serving_client, params, entities):
    passenger_id = params[0]
    df = params[1]
    df['Sex'] = df['Sex'].map( {'female': 0, 'male': 1} ).astype(int)
    return df.loc[df["PassengerId"] == passenger_id]["Sex"].values[0]

At serving time, you can request on-demand features similar to pre-calculated features and alongside pre-calculated features.

features = client.features([("fpf", "quickstart"), ("gender_category", "quickstart")], {"passenger": "1"}, params=["1", df])

​
Wait

The wait() function can be used to wait until a feature table is fully materialized before attempting to lookup the feature. It has an optional timeout in second to wait until for the materialization completion.

fpf = client.features([("fpf", "quickstart")], {"passenger": "1"}).wait(timeout=60)

​
Serving for Training

When a training set is defined, it is materialized into the offline store associated with the definition. The label from the training set is zipped with the features to form a point-in-time correct dataset. Once that’s complete we can initialize a ServingClient and loop through the our dataset

import featureform as ff

client = ff.ServingClient(host)
dataset = client.training_set(name, variant)
for row in dataset:
    # print(row.features(), row.label())
    # Train Model

​
Dataset API

The Dataset API takes inspiration from Tensorflow’s Dataset API, and both can be used very similarly.

​
Repeat

import featureform as ff

client = ff.ServingClient(host)
for row in client.training_set(name, variant).repeat(5):
    # Run through 5 epochs of the dataset

​
Shuffle

import featureform as ff

client = ff.ServingClient(host)
buffer_size = 1000
for row in client.training_set(name, variant).shuffle(buffer_size):
    # Run through a shuffled dataset

​
Batch

import featureform as ff

client = ff.ServingClient(host)
for batch in client.training_set(name, variant).batch(8):
    # Run batches of features and labels

​
Wait

The wait() function can be used to wait until the specified training set is ready to be served before attempting to call it. It has an optional timeout in seconds to wait for the training set to be complete.

import featureform as ff

client = ff.ServingClient(host)
for row in client.training_set(name, variant).wait(timeout=60):
    # Run batches of features and labels

​
Putting it all together

The commands above can be chained together.

import featureform as ff

client = ff.ServingClient(host)
dataset = client.training_set(name, variant).repeat(5).shuffle(1000).batch(64)
for feature_batch, label_batch in dataset:
    # Run through a shuffled dataset 5x in batches of 64

​
Model Registration

We can optionally include a model name at the time of serving features or training sets to create a logical grouping of models and their dependencies.

fpf = client.features([("fpf", "quickstart")], {"passenger": "1"}, model="passengers_random_forest")

Featureform tracks which features and training sets are associated with models registered at serving time. Both features and training_set methods accept the same optional model param, which is typically a string. We can see a list of models in the dashboard and drill down into a specific model to see the features and/or training sets that are associated with it.