Training sets and feature views in Featureform

Build training sets and materialize feature views to Redis with Featureform.

Training sets and feature views are where Featureform turns definitions into outputs that model builders and applications can use.

• training sets support model development
• feature views support online inference

Register metadata first

In the current workflow, client.apply() registers metadata for providers, datasets, transformations, entities, features, and labels:

client.apply()

This is a separate step from creating the Redis-backed serving surface.

Training sets

Register a training set by combining a label with the features you want to train on:

fraud_training_set = client.register_training_set(
    name="fraud_training_set",
    label=User.is_fraud,
    features=[
        User.avg_transaction_amount,
        User.transaction_count[timedelta(days=7)],
    ],
)

After registration, iterate over the training set or convert it to a data frame for analysis:

training_df = client.dataframe(fraud_training_set)

Feature views

Materialize a feature view to Redis when you need online serving:

client.materialize_feature_view(
    view_name="user_features_view",
    inference_store=redis,
    features=[
        User.avg_transaction_amount,
        User.transaction_count[timedelta(days=7)],
        User.transaction_count[timedelta(days=30)],
    ],
)

Use feature views to group the online features that an application or model needs at inference time.

Serve feature values

result = client.serve_feature_view(
    view="user_features_view",
    entity_ids=["user_1", "user_2"],
)

The result contains the latest materialized values stored in Redis for those entity IDs.

Common point of confusion

Keep these two steps separate in your mental model:

• client.apply() registers definitions and lineage
• client.materialize_feature_view(...) creates the serving surface in Redis

That distinction matters when you troubleshoot missing online data. A successful metadata registration does not, by itself, mean the feature view is ready for serving.

Best practices

• Use training sets for repeatable model development workflows.
• Keep online feature views focused on the features needed by one serving surface.
• Treat Redis as the online system of record for materialized feature views, not the source of offline feature computation.

What to read next

• Streaming features