Welcome to the Feast quickstart! This quickstart is intended to get you up and running with Feast in your local environment. It covers the following workflows:

1. Setting up Feast

2. Registering features

3. Constructing training datasets from offline data

4. Materializing feature data to the online feature store

5. Fetching feature vectors for inference

The quickstart uses some example data about a ride-hailing app to walk through Feast. Let's get into it!

A Feast installation includes a Python SDK and a CLI. Feast can be installed from `pip`:

You can test your installation by running`feast version` from your command line:

We can bootstrap a feature repository using the `feast init` command:

This command generates two files. Let's take a look at `feature_store.yaml`:

This file defines how the feature store is configured to run. The most important option here is `provider`, which specifies the environment that Feast will run in. We've initialized `provider=local`, indicating that Feast will run the feature store on our local machine. See [Repository Config](reference/repository-config.md) for more details.

Next, take a look at `example.py`. This file defines some example features:

This file defines three objects:

* A `DataSource`, which is a pointer to persistent feature data. In this example, we're using a `FileSource`, which points to a set of parquet files on our local machine.

* An `Entity`, which is a metadata object that is used to organize and join features. In this example, our entity is `driver_id`, indicating that our features are modeling attributes of drivers.

* A `FeatureView`, which defines a group of features. In this example, our features are statistics about drivers, like their conversion rate and average daily trips.

Feature definitions in Feast work similarly to Terraform: local definitions don't actually affect what's running in production until we explicitly register them with Feast. At this point, we have a set of feature definitions, but we haven't registered them with Feast yet.

We can register our features by running `feast apply` from the CLI.

After this command completes, our features have been registered to Feast, and they're now ready for offline retrieval and materialization.

Feast generates point-in-time accurate training data. In our example, we are using statistics about drivers to predict the likelihood of a booking completing. When we generate training data, we want to know what the features of the drivers were _at the time of prediction_.

Generating training datasets is a workflow best done from an interactive computing environment, like a Jupyter notebook. You can start a Jupyter notebook by running `jupyter notebook` from the command line. Then, run the following code to generate an _entity DataFrame_:

This DataFrame represents the entity keys and timestamps that we want feature values as of. We can pass this Entity DataFrame into Feast, and Feast will fetch point-in-time correct features for each row:

This DataFrame contains all the necessary signals needed to train a model, excluding labels, which are typically managed outside of Feast. Before you can train a model, you'll need to join on labels from external systems.

We just used Feast to generate Using the `local` provider, the online store is a SQLite database. To materialize features, run the following command from the CLI:

We've just populated the online store with the most up-to-date features from the offline store. Our feature values are now ready for real-time fetching.

After we materialize our features, we can use the `store.get_online_features` to fetch the _latest_ feature values for real-time inference. Run the following code in your notebook to fetch online features:

This feature vector can be used for real-time inference, for example, in a model-serving microservice.

This quickstart covered the essential workflows of using Feast in your local environment. The next step is to set `provider="gcp"` in your `feature_store.yaml` file and push your work to production deployment. You can also use the `feast init -t gcp` command in the CLI to initialize a feature repository with example features in the GCP environment.

* See [Create a feature repository](how-to-guides/create-a-feature-repository.md) for more information on the workflows we covered.

* Join our[ Slack group](https://slack.com) to talk to other Feast users and the maintainers!