Skip to content
CSA Loom — the Microsoft Fabric experience for Azure tenants where Fabric isn't yet available: lakehouses, warehouses, notebooks, semantic models, Activator rules, Data Agents, across Commercial, GCC, GCC-High, and DoD IL5

Tutorial: Feature table editor

CSA Loom feature-table editor — a first-class Feature Store surface: Unity Catalog feature tables offline, Lakebase / pgvector online, with point-in-time joins and feature lookup at inference. One-for-one with the Databricks Feature Engineering experience, Loom-themed and sovereign — the Azure-native default is Databricks UC; Gov uses OSS Unity Catalog + Azure Database for PostgreSQL. No Microsoft Fabric.

What it is

A feature store solves two hard problems: training/serving skew and feature reuse. This editor gives you both halves:

  • Offline — a versioned feature table (Delta on Unity Catalog, or PostgreSQL on the sovereign path) with entity keys and a timestamp key, joined onto training spines as-of the label time so a model never sees the future.
  • Online — the same features published to a low-latency store (Lakebase / pgvector) and looked up by entity key at inference time, then merged into the scoring payload for a model-serving endpoint.

Four tabs: Overview, Define, Point-in-time join, Online serving (the last two unlock once the table is defined).

When to use it

  • Multiple models need the same features and you want one definition, one refresh, one lineage.
  • You need training data assembled correctly as-of each label's timestamp.
  • You need those same features available at inference in milliseconds.

Step-by-step in Loom

  1. Create the item. + New item → Feature table, then Create feature table item. The header badges state the active backend — Unity Catalog feature tables (Azure-native default) or OSS Unity Catalog + PostgreSQL (sovereign / Gov path) — plus Online: Lakebase / pgvector, and a defined badge once a spec exists. The right Details panel summarizes the table, offline backend, entity keys, timestamp key, feature count, and online table.
  2. Define the feature table. On Define fill in:
  3. Catalog / Schema / Table — pre-filled with the deployment's defaults where available; together they compose the three-part full name.
  4. Entity (primary) keys — comma-separated (for example customer_id).
  5. Timestamp key — the event-time column (for example event_ts).
  6. Feature columns — Add feature rows, each a name plus a type from DOUBLE, FLOAT, BIGINT, INT, STRING, BOOLEAN, TIMESTAMP, DATE.

Create feature table creates the real offline table (Delta or PostgreSQL) and the online table, then persists the spec. The button becomes Update feature table afterwards. 3. Build a training set with a point-in-time join. On Point-in-time join: - Spine / training table — your labels table. - Spine entity keys — comma-separated, aligned to the feature keys (pre-filled from the spec). - Spine timestamp key — the label time. - Carry columns — the label columns to carry through. - Row limit — defaults to 1000.

Preview SQL shows the generated AS-OF join without running it. Run join executes it and returns real, type-badged rows with a row count and execution time. 4. Publish to the online store. On Online serving, Publish latest features materializes the current offline rows into the online table and reports how many entity rows were published (or states honestly that there are no offline rows yet and the online table is ready). 5. Look up features and score a model. Still on Online serving: - Serving endpoint — the model-serving endpoint name (for example fraud-scorer). - One input per entity key — the identity to look up. - Scoring payload (JSON) — the request body; the looked-up features are merged in before it is sent.

Look up + invoke performs the online lookup and calls the endpoint, returning the resolved features, the HTTP status, the end-to-end latency, and the model's response body. 6. Reload. Reload in the ribbon re-reads the spec, backend, and gate state.

The Azure backend it rides on

  • Offline store (default): Databricks Unity Catalog feature tables on Delta.
  • Offline store (sovereign / Gov): OSS Unity Catalog + Azure Database for PostgreSQL.
  • Online store: Lakebase / pgvector for low-latency key lookups.
  • Serving: a Loom model-serving endpoint (see editor-model-serving-endpoint.md).
  • Routes: GET/POST /api/items/feature-table/<id>, …/pit-join, …/online, …/serve — every control calls a real BFF route; there are no mocks and no dead buttons.

Honest gates

Condition What you see Exact remediation
Offline backend not configured The shared Fix-it gate (svc-feature-store) with the exact missing variable and a hint; the Define form's inputs are disabled but the surface still renders — no red banner on a fresh item Set the env var the gate names, using its inline Fix it wizard
Online store not configured A separate online gate; Publish and Look up + invoke are disabled Set the online-store variable the gate names
Table not defined yet Point-in-time join and Online serving tabs are disabled Define the feature table first
Nothing published yet Publish reports "No offline rows to publish yet" and names the ready online table Land offline feature rows, then publish again

No Fabric required

Databricks UC (or OSS UC + Azure Database for PostgreSQL) plus Lakebase / pgvector. No Fabric capacity, workspace, OneLake path, or Power BI workspace is used on any path.

Learn more

  • Model serving endpoint editor tutorial: editor-model-serving-endpoint.md
  • ML model editor tutorial: editor-ml-model.md
  • Parity source: docs/fiab/parity/feature-store.md