A hands-on example pipeline that combines Django + django-mongodb-backend, Feast, and MongoDB to detect potentially fraudulent transactions.
- Transactions are stored in MongoDB through the Django ORM (
transactionsapp). - Feast defines and manages
transaction_features(amount,category,location), using a local Parquet file as the offline store and MongoDB as the online store. - A simple z-score model (
training/train_model.py) is trained on historical features pulled from Feast. - Features are materialized into MongoDB via
feast materialize, and a Django view (transactions/views.py) scores transactions in real time by looking up each user's latest materialized features.
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# Create a .env file with SECRET_KEY, MONGODB_URI, and DB_NAME (see tutorial.md)
# The `feast` CLI reads MONGODB_URI and DB_NAME from the environment.
set -a; source .env; set +a
python data/generate_transactions.py
python manage.py migrate
python manage.py load_transactions
cd feature_repo/feature_repo
feast apply
# The end date must be after the newest event_timestamp in your data.
feast materialize 2020-01-01T00:00:00 2026-09-01T00:00:00
cd ../..
# Writes training/model_params.json, which the scoring view reads.
cd training && python train_model.py && cd ..
python manage.py runserverThen request a fraud score:
curl http://127.0.0.1:8000/score/2/{
"user_id": 2,
"amount": 28.079999923706055,
"category": "entertainment",
"location": "London",
"z_score": -0.476,
"is_potentially_fraudulent": false
}
