Feature Views are now generally available on Databricks, per the October 2026 release notes dated 12 October 2026, covering batch, streaming and request-time data. Databricks says they will be enabled for compliance security profile workspaces in mid-November 2026. Databricks had announced the Public Preview on its blog, describing a framework where you author a feature once and the platform carries it from experimentation through to real-time serving.
Our read on why this matters is the code it removes. It's a much needed feature that will help teams move forward with building features by reducing the code and focusing on the things that bring value to the business. Feature pipelines are where data teams accumulate the most duplicated logic. The same window, the same join key and the same lookback get written once for the training notebook and again for the serving path. One definition that both sides read is worth more than any individual capability in the list.
What you need to run it #
The requirements are short, and they're the part to check before anyone plans a sprint. You need serverless compute or a classic cluster on a Databricks Runtime ML version that supports Feature Views, and the client package installed in each notebook run:
%pip install databricks-feature-engineering
dbutils.library.restartPython()The client release notes describe create_feature(), which validates CustomUDF input bindings against the Unity Catalog function signature when the feature is created. create_training_set() is where windowed features' start_time gating decides which training rows are kept. materialize_features() accepts tags and budget_policy_id for cost tracking and attribution.
Pin the client, and read its release notes #
The Feature Store release notes, last updated 1 October 2026, are where the sharp edges live. Client 0.18.0 changed windowed feature behaviour so start_time is always enforced: training rows whose window end falls before a feature's start_time are excluded, so that offline training sets match the values served online. That's a correctness fix for training-serving skew, and it also means a client bump can legitimately change the rows in a training set. The same release made aggregation features reject DATE timeseries columns at creation, so use a TIMESTAMP. Client 0.18.1 added support for feature names longer than 64 characters in Lakebase online serving.
The stage labels haven't all caught up. SawtoothWindow is listed as Beta, and the Feature Views API reference still carries a Public Preview banner. The release note is the stage of record, so cite it in anything a reviewer reads.
What to do next #
Check the ML runtime version first; a runtime too old for Feature Views is the blocker, not the feature. Then take one feature that's computed twice, once in a training notebook and once in a serving path, and redefine it as a single feature, setting start_time where it is windowed to control the cutoff. That's the duplicate this is built to delete, and it's where skew usually hides. Databricks says compliance security profile workspaces will have Feature Views enabled in mid-November 2026, so run that review before it lands.
The lifecycle this plugs into is in MLflow and the Model Lifecycle, and the online read path is in Databricks Apps.
Questions people ask #
Are Databricks Feature Views generally available? #
Yes. Databricks' October 2026 release notes state that Feature Views are generally available as of 12 October 2026, for batch, streaming and request-time data. Some related API reference pages still display a Public Preview banner.
What do I need to use Feature Views? #
Serverless compute or a classic compute cluster on a Databricks Runtime ML version that supports Feature Views, plus the databricks-feature-engineering package installed and followed by dbutils.library.restartPython() in each notebook run.
What sources can a Feature View use? #
Batch, streaming and request-time data; the client release notes mention Delta sources. The client supports windowed aggregations, CustomUDF features and features composed from other features via FeatureViewSource.