Feature Tables in Unity Catalog

What it is

A feature table stores precomputed model inputs, one row per entity (customer, product, device), keyed by a primary key. In Unity Catalog a feature table is simply a Delta table with a primary key constraint, governed like any other table.

The exam says "feature store table". Today the product is called Feature Engineering in Unity Catalog, installed as the databricks-feature-engineering package and used through FeatureEngineeringClient. The older workspace Feature Store and its FeatureStoreClient are legacy and apply only to workspaces not enabled for Unity Catalog.

How it works

Account level versus workspace level

A Unity Catalog metastore sits at the account level and can be attached to many workspaces. A workspace feature store belongs to one workspace.

The headline benefits are reuse across workspaces, central governance, and lineage. A team in the production workspace can read features that the development team built, without copying data.

Creating a feature table

You need a primary key. Create the table from a DataFrame (schema and data in one step) or from a schema only, then write data later.

from databricks.feature_engineering import FeatureEngineeringClient
fe = FeatureEngineeringClient()
fe.create_table(
    name="ml.retail.customer_features",
    primary_keys=["customer_id"],
    df=customer_features_df,
    description="Customer purchase features")

Passing schema=customer_features_df.schema instead of df creates an empty table with that structure. For time series features, add the time column to primary_keys and name it in timeseries_columns, which enables point in time lookups later.

You can also create one in SQL, because any Delta table with a primary key is a feature table. The primary key columns must be NOT NULL.

CREATE TABLE ml.retail.customer_features (
  customer_id BIGINT NOT NULL,
  orders_30d INT,
  CONSTRAINT customer_pk PRIMARY KEY (customer_id));

Writing data to a feature table

write_table writes a DataFrame into an existing feature table.

fe.write_table(
    name="ml.retail.customer_features",
    df=new_features_df,
    mode="merge")

With mode="merge", rows whose primary key already exists are updated, new keys are inserted, and rows not in the DataFrame are left unchanged. This is an upsert, which makes daily refresh jobs safe to rerun.

Worked example

The table holds customers 1, 2 and 3. A refresh DataFrame contains customers 2 and 4. After a merge, customer 2 is updated, customer 4 is added, and customers 1 and 3 keep their old values, so the table has four rows.

How the exam asks about this

• "Features built in one workspace must be reused by teams in two other workspaces with central access control." Create them as Unity Catalog feature tables, which every workspace on the metastore can use under UC grants.
• "Update today's feature values for existing customers and add new customers, keeping all other rows." Call write_table with mode="merge".
• "Which argument is required to create a feature table?" The primary_keys (with the three-level name), plus either a df or a schema.