By the end of this guide, you will have a clear understanding of:<\/p>\n
- \n
- What is Feast?<\/li>\n
- Key Feast components<\/li>\n
- Feast Operator setup on Kubernetes<\/li>\n
- Deploy Feast using CRD and configure offline and online stores.<\/li>\n
- Use a simple Python script to verify feature serving<\/li>\n<\/ul>\n
Lets get started.<\/p>\n
What is Feast?<\/h2>\n
Feast is an open source feature store used by teams to store, manage, and serve ML features for both training and inference. <\/p>\n
Feast acts as a centralized storage for features that will be used by different teams. The following image illustrates the Feast architecture on Kubernetes<\/a>. <\/p>\n
![\"Architecture](\"https:\/\/storage.ghost.io\/c\/5f\/2f\/5f2f4d20-2abf-4534-8d40-7aa233aedd43\/content\/images\/2026\/04\/image-112.png\")
<\/figure>\nThe following are the key components of Feast.<\/p>\n
- \n
- PostgreSQL Feature Registry<\/strong>: It stores the definitions (schema) of the features. Eg,. age, name, etc.<\/li>\n
- Redis Online store<\/strong>: It is used to reduce the latency during inferencing.<\/li>\n
- Offline store<\/strong>: It is used to store historical features used for model training.<\/li>\n
- Feast pod<\/strong>: It serves latest features to ML model by reading from the registry and the online store.<\/li>\n
- Feast CLI<\/strong>: It is a tool used to connect Feast and to run Feast commands. It connects to feature registry, offline Store and online Store<\/li>\n<\/ul>\n
Lets get started with the setup.<\/p>\n
Setup Prerequisites<\/h2>\n
Following are the prerequisites to follow this guide.<\/p>\n
- \n
- Kubernetes cluster<\/a><\/li>\n
- kubectl<\/a><\/li>\n
- Python 3+<\/li>\n
- Pip<\/li>\n<\/ul>\n
GitHub Repository<\/h2>\n
First, fork and clone the MLOPs Github repository. <\/a><\/p>\n
https:\/\/github.com\/techiescamp\/mlops-for-devops.git<\/code><\/pre>\nCD into the
mlops-for-devops<\/code> directory.<\/p>\n\ncd mlops-for-devops<\/code><\/pre>\n\nNote:<\/strong><\/b> I have already pushed the required manifest file for PostgreSQL. Redis and Feast CR inside theplatform-tools\/feast<\/code> folder of the repository that you cloned before.<\/div>\n<\/div>\nThe files we are going to use in this guide are shown below.<\/p>\n
.\n\u251c\u2500\u2500 phase-1-local-dev\n\u2502 \u2514\u2500\u2500 datasets\n\u2502 \u2514\u2500\u2500 employee_attrition.csv\n\u251c\u2500\u2500 phase-2-enterprise-setup\n\u2502 \u2514\u2500\u2500 feast\n\u2502 \u251c\u2500\u2500 csv_to_parquet.py\n\u2502 \u251c\u2500\u2500 feature_store.yaml\n\u2502 \u251c\u2500\u2500 feature_definitions.py\n\u2502 \u2514\u2500\u2500 smoke_test.py\n\u2514\u2500\u2500 platform-tools\n \u2514\u2500\u2500 feast\n \u251c\u2500\u2500 feast.yaml\n \u251c\u2500\u2500 psql.yaml\n \u2514\u2500\u2500 redis.yaml\n<\/code><\/pre>\nSetup Feast on a Kubernetes Cluster<\/h2>\n
The first step is to setup Feast on a Kubernetes cluster<\/a> using the Feast Operator.<\/a><\/p>\n
Here is what we are going to do.<\/p>\n
- \n
- Setup feast operator<\/li>\n
- Deploy Redis and PostgreSQL<\/a> for the registry and Online store<\/li>\n
- Uses
FeatureStore<\/code><\/strong> Custom Resource<\/a> to create Feast feature store<\/li>\n<\/ul>\n\n\ud83d\udca1<\/div>\nWe will expose Redis, PostgreSQL, and Feast as NodePort, because we need to access it from local to store and get features.<\/p>\nIn actual setup, use Ingress<\/a> or Gateway API <\/a>to expose Feast externally.<\/div>\n<\/div>\n
Lets get started.<\/p>\n
Step1: Install Feast Operator<\/h3>\n
Use the following command to deploy the Feast Operator. It will deploy the operator on the
feast-operator-system<\/code> namespace.<\/p>\nkubectl apply --server-side --force-conflicts -f https:\/\/raw.githubusercontent.com\/feast-dev\/feast\/refs\/heads\/stable\/infra\/feast-operator\/dist\/install.yaml<\/code><\/pre>\nRun the following command to verify if the operator is running and CRD is created.<\/p>\n
$ kubectl get pods -n feast-operator-system\n\n$ kubectl get crd featurestores.feast.dev<\/code><\/pre>\nYou should see the following output.<\/p>\n
NAME READY STATUS RESTARTS AGE\n\nfeast-operator-controller-manager-482tt 1\/1 Running 0 38s\n\nNAME CREATED AT\n\nfeaturestores.feast.dev 2026-01-08T10:32:38Z<\/code><\/pre>\nStep 2: Create a Namespace for Feast Deployment<\/h3>\n
We will be deploying the Feast pods on a dedicated namespace named
feast<\/code><\/strong>. Create it using the following command.<\/p>\n$ kubectl create namespace feast<\/code><\/pre>\nStep 3: Setup PostgreSQL for Feast Registry<\/h3>\n
We need to set up PostgreSQL as a registry for Feast. It stores the metadata of Feast.<\/p>\n
Assuming you are in the
root<\/strong><\/code> directory, run the following command to move into theplatform-tools\/feast<\/code> folder.<\/p>\ncd platform-tools\/feast<\/code><\/pre>\nYou will find the following files.<\/p>\n
.\n\u251c\u2500\u2500 feast.yaml\n\u251c\u2500\u2500 psql.yaml\n\u2514\u2500\u2500 redis.yaml<\/code><\/pre>\nThe
psql.yaml<\/code><\/strong> manifest, deploys a kubernetessecret<\/code><\/strong> that contains the PostgreSQL username, password, and database name as shown below. (feast<\/strong> as default)<\/p>\n![\"modify](\"https:\/\/storage.ghost.io\/c\/5f\/2f\/5f2f4d20-2abf-4534-8d40-7aa233aedd43\/content\/images\/2026\/04\/image-12.png\")
<\/figure>\nIt also contains manifest to deploy PostgreSQL as a StatefulSet<\/a> and Service to expose PostgreSQL as Nodeport.<\/p>\n
Now, apply the PostgreSQL manifest using the following command.<\/p>\n
$ kubectl apply -f psql.yaml<\/code><\/pre>\nThen run the following command to ensure the postgres pod<\/a> is up and running.<\/p>\n
$ kubectl get po -n feast\n\nNAME READY STATUS RESTARTS AGE\n\npostgres-0 1\/1 Running 0 34s<\/code><\/pre>\nStep 4: Setup Redis for Feast Online Store<\/h3>\n
Next, we need to deploy Redis Statefulset<\/strong>. It will be the online store for Feast.<\/strong><\/p>\n
The
redis.yaml<\/code> manifest creates asecret<\/code> that contains the Redis password, Redis as StatefulSet, and a Service to expose Redis as Nodeport.<\/p>\n![\"\"](\"https:\/\/storage.ghost.io\/c\/5f\/2f\/5f2f4d20-2abf-4534-8d40-7aa233aedd43\/content\/images\/2026\/04\/image-14.png\")
<\/figure>\nNow, apply the manifest using the following command.<\/p>\n
$ kubectl apply -f redis.yaml<\/code><\/pre>\nThen run the following command to ensure the redis pod is up and running.<\/p>\n
$ kubectl get po -n feast\n\nNAME READY STATUS RESTARTS AGE\n\npostgres-0 1\/1 Running 0 1m\nredis-0 1\/1 Running 0 8s<\/code><\/pre>\nYou can see a Redis pod along with the PostgreSQL pod.<\/p>\n
\n\ud83d\udca1<\/div>\nIn this setup, we will be using a CSV file directory as offline store<\/strong><\/b>. For production environment, the offline store would be usually on BigQuery, Snowflake, Redshift, and others<\/a>.<\/div>\n<\/div>\nStep 5: Deploy Feast Feature Store<\/h3>\n
Now its time to deploy the Feast feature store using the
FeatureStore<\/code><\/strong> CR. It tells the Feast operator to deploy the Feast pod, its service, and configure the backend storage.<\/p>\nThe
feast.yaml<\/code> manifest will create a secret that contains the connection string of Redis and PostgreSQL<\/strong>, along with the credentials to authenticate them.<\/p>\nAdditionally, it will also create an extra service to expose it as a NodePort.<\/p>\n
Apply the manifest using the following command.<\/p>\n
$ kubectl apply -f feast.yaml<\/code><\/pre>\nThen run the following command to check if the Feast deployment<\/a> is up and running.<\/p>\n
$ kubectl get po -n feast\n\nNAME READY STATUS RESTARTS AGE\n\nfeast-feast-tbn5k 1\/1 Running 0 30s\npostgres-0 1\/1 Running 0 3m\nredis-0 1\/1 Running 0 2m<\/code><\/pre>\nYou can see that all three pods are running.<\/p>\n
Feature Repository<\/h2>\n
The
phase-2-enterprise-setup\/feast<\/code><\/strong> directory which contains all the feast Python files to create, store, and retrieve features. It is called the feature respository.<\/strong><\/p>\ncd ..\/..\/phase-2-enterprise-setup\/feast\/<\/code><\/pre>\nHere is the structure of our feature repository.<\/p>\n
.\n\u251c\u2500\u2500 csv_to_parquet.py\n\u251c\u2500\u2500 feature_store.yaml\n\u251c\u2500\u2500 feature_definitions.py\n\u251c\u2500\u2500 smoke_test.py\n\u2514\u2500\u2500.feastignore<\/code><\/pre>\nIn this:<\/p>\n
- \n
feature_store.yaml<\/code><\/strong> – It is the configuration file that Feast CLI uses to read the Redis and PostgreSQL connection strings. It should be present in root of a feature repository<\/li>\ncsv_to_parquet.py<\/code><\/strong> – Contains logic to process the CSV file and convert it to a Parquet file.<\/li>\nfeature_definitions.py<\/code><\/strong>– It is the schema\/contract<\/strong> for the feature store (feature registry definition file) . Meaning, it defines what features exist, where the data comes from, and how they should be served<\/li>\nsmoke_test.py<\/code><\/strong> – It is a smoke-test script that hits the Feast online feature store HTTP API and verifies that real-time feature retrieval is working end-to-end.<\/li>\n.feastignore<\/code><\/strong> – Contains file and folder names that should be ignored by feast apply.<\/li>\n<\/ul>\nStore Features into Feast feature store<\/h2>\n
Now, our setup is ready. Now lets look at how to store features into Feast feature store. Both offline and online store.<\/p>\n
Lets get started.<\/p>\n
Step 1: Setup Feast CLI<\/h3>\n
In the local environment, we need to install Feast CLI along with PostgreSQL and Redis packages.<\/p>\n
Run the following command to create a isolated virtual environment and install the Feast CLI with PostgreSQL and Redis packages.<\/p>\n
python3.11 -m venv env\n\nsource env\/bin\/activate\n\npip install --upgrade pip setuptools wheel\n\npip install 'feast[postgres,redis]'<\/code><\/pre>\nStep 2: Update NodePort IP to feature_store.yaml<\/h3>\n
Since we need to run the feast commands from our local workstations, we have exposed Feast as NodePort. So you need to update one of the cluster node IP in the
feature_store.yaml<\/code><\/strong> file.<\/p>\nTo get the node IP, run the following command.<\/p>\n
kubectl get no -o wide<\/code><\/pre>\n\n\ud83d\udca1<\/div>\nIdeally the feast endpoint would be a kubernetes service url if feast is running in the same cluster or a ingress<\/a> endpoint if it is in a different cluster.<\/div>\n<\/div>\nStep 3: Setup Offline Source (Convert CSV to Parquet)<\/h3>\n
For model training, an offline store is used. For file based feast offline store<\/strong>, we need to convert the CSV to Parquet format.<\/strong><\/p>\n
\n\ud83d\udca1<\/div>\nFeast’s file-based offline store (local\/S3\/GCS) uses Parquet<\/strong><\/b><\/a> (A column-oriented data file format) & Delta<\/a> as its source format. <\/p>\nFor other data warehouse-based offline stores like BigQuery and Redshift, Feast interacts with native tables and SQL queries.<\/p><\/div>\n<\/div>\n
The
csv_to_parquet.py<\/code><\/strong> has the code to generate the Parquet file by processing the data on theemployee_attrition.csv<\/code><\/strong> file. <\/p>\n![\"code](\"https:\/\/storage.ghost.io\/c\/5f\/2f\/5f2f4d20-2abf-4534-8d40-7aa233aedd43\/content\/images\/2026\/04\/image-116.png\")
<\/figure>\nHere is high level overview of what the script does.<\/p>\n
- \n
- Reads
employee_attrition.csv<\/code> from yourphase-1-local-dev\/datasets\/<\/code><\/li>\n- Since ML models need numbers, it converts
Yes\/No<\/code> columns (Overtime, Remote Work, etc.) to1\/0<\/code><\/li>\n- Feast expects
snake_case<\/strong><\/code> column names. So it convertsMonthly Income\"<\/code> to\"monthly_income<\/code> etc<\/li>\n- Add Feast-required timestamp column like
event_timestamp<\/code><\/strong> &created<\/code><\/strong><\/li>\n- Writes the final output to
data\/employee_attrition_features.parquet<\/code> . This file becomes the offline data source<\/strong> that Feast reads from.<\/li>\n<\/ul>\nRun the Python Script
csv_to_parquet.py<\/code><\/strong> using the following command.<\/p>\npython csv_to_parquet.py<\/code><\/pre>\nNow, if you check the
\/phase-2-enterprise-setup\/feast\/data<\/code> folder, you can see aemployee_attrition_features.parquet<\/code> file created.<\/p>\n\n\ud83d\udca1<\/div>\nNow we are using the offline store in local environment under the\/data<\/strong><\/code><\/b> folder. It is perfect for local development. <\/p>\nDuring actual training, the feature data (Parquet file) is generated using data pipelines ( Kubeflow<\/a> or Airflow<\/a>) and stored on services like S3, GCS etc. <\/div>\n<\/div>\n
Step 4: Register Features to Feast<\/h3>\n
I already have a Parquet file. So why do I need to register features in Feast?<\/em><\/p>\n
Well, Parquet file only stores data. Feast needs a structured definition (schema, entity, source, and usage) to serve those features consistently across training and inference.<\/p>\n
Registering features involves two commands:<\/p>\n
- Since ML models need numbers, it converts
- Uses
- kubectl<\/a><\/li>\n
- Redis Online store<\/strong>: It is used to reduce the latency during inferencing.<\/li>\n
- PostgreSQL Feature Registry<\/strong>: It stores the definitions (schema) of the features. Eg,. age, name, etc.<\/li>\n
![\"image](\"https:\/\/storage.ghost.io\/c\/5f\/2f\/5f2f4d20-2abf-4534-8d40-7aa233aedd43\/content\/images\/2026\/04\/image-114.png\")
<\/figure>\n