Real-time credit scoring application with Feast & ScyllaDB

This repository is based on this existing Feast sample application.

This sample project is a real-time credit scoring application example that shows you how to set up Feast with ScyllaDB Cloud as an online store and parquet files as offline store.

What is a feature store?

Feature store is a central data store to power operational machine learning models. They help you store transformed feature values in a scalable and performant database. Real-time inference requires features to be returned to applications with low latency at scale. This is where ScyllaDB can play a crucial role in your machine learning infrastructure.

Why should you consider ScyllaDB as feature store?

ScyllaDB is a real-time NoSQL database that is best suited for feature store use cases where you require low latency (e.g. model serving), high throughout (e.g. training) and need peta-byte scalability.

• Low-latency: ScyllaDB can provide <1 ms P99 latency. For real-time machine learning apps, an online feature store is required to meet strict latency requirements. ScyllaDB is an excellent choice for an online store (Read how Medium is using ScyllaDB as a feature store.)
• High-throughput: Training requires querying huge amounts of data and processing large datasets with possibly millions of operations per second - something that ScyllaDB excels at.
• Large-scale: ScyllaDB can handle petabytes of data while still keeping latency low and predictable.
• High availability: ScyllaDB is a highly available database. With its distributed architecture, ScyllaDB keeps your feature store database always up and running.
• Easy to migration: ScyllaDB is compatible with DynamoDB API and Cassandra which means it's simple to migrate over from legacy solutions.
• Integration with Feast: ScyllaDB is compatible with DynamoDB API and Cassandra which means it's simple to migrate over from legacy solutions.

Requirements

• Python 3
• ScyllaDB Cloud account

Get started

Clone the repository:

git clone https://github.com/zseta/scylladb-feast
cd scylladb-feast

Project details:

• The primary training dataset is in loan_table.parquet. This file contains historic loan data with accompanying features. The dataset also contains a target variable, namely whether a user has defaulted on their loan.
• Feast is used during training to enrich the loan table with zipcode and credit history features from other parquet files.
• Feast is also used to serve the latest zipcode and credit history features for online credit scoring using ScyllaDB Cloud.

In this tutorial, you'll do the following steps:

1. Create a new ScyllaDB Cloud cluster
2. Create a new keyspace for the feature store
3. Install dependencies and configure Feast
4. Deploy and test the feature store

Create ScyllaDB Cloud cluster

Go to ScyllaDB Cloud and create a new cluster (either "Free Tial" or "Dedicated VM"). You can use the smallest available machine for this sample app (t3.micro)

Create new keyspace

Install CQLSH command line tool and connect to the ScyllaDB cluster:

pip install cqlsh
cqlsh <SCYLLA-CLOUD-HOST> -u scylla -p <PASSWORD>

You can get the host address, username, and password from your ScyllaDB Cloud dashboard:

Create a new keyspace called feast in ScyllaDB (this keyspace will be populated by Feast):

CREATE KEYSPACE feast WITH replication = {'class': 'NetworkTopologyStrategy', 'replication_factor': '3'} ;

Install dependencies and configure Feast

Create a new Python environment and install dependencies:

virtualenv env && source env/bin/activate
pip install -r requirements.txt

Next, configure ScyllaDB Cloud as the online store for Feast.

Tip

ScyllaDB is compatible with Apache Cassandra so you can use the Feast Cassandra connector with ScyllaDB

Open the feature_store.yaml file and add the host addresses, username (scylla), and password for ScyllaDB:

project: repo
# By default, the registry is a file (but can be turned into a more scalable SQL-backed registry)
registry: data/registry.db
# The provider primarily specifies default offline / online stores & storing the registry in a given cloud
provider: local
online_store:
    type: cassandra
    hosts:
        - x.x.x.x
        - x.x.x.x
        - x.x.x.x
    username: scylla
    password: pass
    keyspace: feast
entity_key_serialization_version: 2

You can get these values from the ScyllaDB Cloud dashboard.

Deploy and test the feature store

Deploy the feature store by running apply from within the feature_repo/ folder

cd feature_repo/
feast apply


Deploying infrastructure for credit_history
Deploying infrastructure for zipcode_features

This commmand also generates a registry.db file in your data/ folder containing Feast configuration information.

Next, load features into the online store using the materialize-incremental command. This command loads the latest feature values from a data source (parquet files, in our case) into the online store (ScyllaDB).

CURRENT_TIME=$(date -u +"%Y-%m-%dT%H:%M:%S")
feast materialize-incremental $CURRENT_TIME

The loading process starts:

 35%|███████████████████▋                                     | 9965/28844 [01:20<02:39, 118.37it/s]

When it's completed, train the model using run.py (in the root folder of the repository)

cd ..
python run.py

The script returns the result of a single loan application:

Loan rejected!

You can also run feast ui to explore your features:

feast ui

Open http://0.0.0.0:8888

Interactive demo

Run the Streamlit app locally:

streamlit run app.py

Go to http://localhost:8501

You can modify the variables on the left sidebar then the app will automatically run the prediction using the model.