ChurnFlow: End-to-End MLOps Pipeline
A production-grade MLOps pipeline for telecom customer churn prediction — integrating FastAPI model serving, Docker Compose multi-service orchestration, and GitHub Actions CI/CD in a single end-to-end system.
- Baseline Model Accuracy
- Predictive Power
- CI/CD
- Fully Automated GitHub Actions Pipeline
ArchitectureMLOps · Full Lifecycle PipelineBuilt WithFastAPI · EvidentlyAI · MLflow · Docker · GitHub Actions · scikit-learn · Python
The Problem
Most ML models never make it to production — and those that do degrade silently without anyone noticing
The gap between a trained ML model and a production ML system is one of the most underestimated challenges in applied machine learning.
A model that performs well in a notebook faces a gauntlet of production realities: it needs a serving endpoint that handles real-time requests, a monitoring layer that detects when input data drifts away from the training distribution, an automated retraining trigger that responds to that drift, an experiment tracking system that maintains a full audit trail of every model version, and a CI/CD pipeline that tests, builds, and deploys changes without manual intervention.
Without all of these components working together, the model either never ships or ships and silently degrades as the real world diverges from its training data — both outcomes that represent a complete failure to deliver ML value.
The Solution
A full MLOps lifecycle system — from training to serving to drift detection to CI/CD — built as a production-ready multi-service platform
ChurnFlow is an end-to-end MLOps pipeline for telecom customer churn prediction that demonstrates the full production ML lifecycle in a single integrated system. A scikit-learn Logistic Regression model is trained on telecom churn data, serialized with joblib, and served via a FastAPI REST endpoint at /predict for real-time inference.
MLflow tracks every experiment — parameters, metrics, and model artifacts — maintaining a complete versioned audit trail. EvidentlyAI monitors incoming prediction data for drift and automatically triggers model retraining when distribution shifts are detected.
A Docker Compose configuration orchestrates three separate containers — API, monitoring, and MLflow — as an isolated, reproducible multi-service environment.
GitHub Actions drives the CI/CD pipeline: on every merge to main, automated tests run via pytest, all three Docker images are built and validated, and the updated system is deployment-ready — all without manual intervention.
Key Outcome
A production-ready MLOps system that demonstrates the full ML engineering lifecycle — real-time serving, drift-triggered retraining, experiment tracking, multi-service containerization, and automated CI/CD — built around a telecom churn prediction model that achieves 80% accuracy and AUC of 84%.
Technical Deep Dive
Architecture & Design
Model Development
Model Serving
Monitoring & Automated Retraining
Containerization · Docker Compose
CI/CD · GitHub Actions
Stage by Stage
- Model Development
Data Pipeline & Training
py trains a scikit-learn Logistic Regression model achieving 80% accuracy and AUC of 84%. The trained model is serialized to the models/ directory via joblib for downstream serving.
- Experiment Tracking
MLflow Model Registry
Every training run is logged to MLflow — parameters, metrics, and model artifacts are stored in the mlruns/ directory and accessible through the MLflow UI at localhost:5000. This creates a complete, versioned audit trail across all training iterations, making it possible to compare runs, roll back to a previous version, and trace any production model to the exact data and code that produced it.
- Model Serving
FastAPI REST Endpoint
app.py exposes the trained model as a REST API at localhost:8000/predict, accepting JSON feature payloads and returning churn probability scores in real time. The endpoint loads the serialized model via joblib and is documented through an auto-generated Swagger UI — making it immediately consumable by downstream applications, dashboards, or internal tooling without any additional integration work.
- Monitoring
EvidentlyAI Drift Detection
py runs EvidentlyAI against incoming prediction data, comparing live feature distributions against the training baseline to detect data drift. html. pkl to the models/ directory, and hot-reloading the serving endpoint without manual intervention.
- Containerization
Docker Compose Multi-Service
Three dedicated Dockerfiles — api.Dockerfile, monitor.Dockerfile, mlflow.Dockerfile — define isolated containers for each service. docker-compose.yml orchestrates all three together, enabling the entire system to be spun up with a single docker compose up --build command. This ensures full environment reproducibility across development, testing, and deployment contexts.
- CI/CD
GitHub Actions Automation
Every push or pull request to main triggers the GitHub Actions pipeline — pytest runs the full test suite, and on a successful pass, all three Docker images are built and validated automatically. Artifacts are stored in both GitHub Actions and MLflow. The result is a system where code changes are tested and containerized without any manual build steps, keeping the deployment path clear of human bottlenecks.
Why It Works
Key Design Decisions
Three isolated containers prevent service coupling
Running the API, monitoring, and MLflow as separate Docker containers rather than a monolithic service means each component can be updated, scaled, or replaced independently. A change to the monitoring logic does not require rebuilding the API container. A new model version does not interrupt the drift detection service. This separation of concerns is a foundational production ML engineering principle that ChurnFlow demonstrates in practice.
Drift detection closes the retraining loop automatically
Without automated drift detection, a deployed model silently degrades as real-world data distribution shifts away from the training data — the most common failure mode for production ML systems. By connecting EvidentlyAI's drift reports directly to retrain.py, ChurnFlow ensures that model degradation triggers a response rather than just a report. The retraining loop closes automatically — new data, new model, new versioned .pkl saved to disk, updated endpoint.
CI/CD eliminates manual deployment as a production bottleneck
Manual deployment steps — building images, running tests — are a common source of deployment errors and a significant drag on iteration speed. By automating the entire sequence through GitHub Actions, ChurnFlow ensures that every code change that passes tests is automatically containerized and deployment-ready. This removes the human bottleneck from the deployment path and makes the system's release cycle as fast as the test suite.
Results & Metrics
What the system delivers
- Model Accuracy
- Baseline Logistic Regression on telecom churn dataset — validated on held-out test set
- Predictive Power
- Area under the ROC curve — strong discriminative power between churn and non-churn customers
- CI/CD
- Fully Automated Pipeline
- GitHub Actions — automated testing, Docker build, and Docker Hub push on every main branch merge
Closed-loop drift detection and automated retraining
EvidentlyAI continuously monitors incoming feature distributions against the training baseline. When drift is detected, retrain.py triggers automatically — retraining the model, versioning it in MLflow, and updating the serving endpoint without any manual intervention. The full retraining loop closes without human involvement, keeping the production model aligned with live data distributions.
Full system up in one command — docker compose up --build
Three dedicated Dockerfiles define isolated, reproducible environments for the API, monitoring, and MLflow services. Docker Compose orchestrates all three together — the entire production stack can be spun up from scratch with a single command in any environment. This makes the system portable across development, staging, and production without environment-specific configuration changes.
Zero-touch deployment on every successful merge
Every push to main triggers the GitHub Actions pipeline — pytest runs the full test suite, and on a successful pass all three Docker images are built and pushed to Docker Hub automatically. No manual build steps, no manual pushes, no deployment bottlenecks. Code changes move from commit to deployment-ready container without any developer intervention beyond the merge itself.
Complete versioned audit trail across every model iteration
MLflow logs parameters, metrics, and model artifacts for every training and retraining run — creating a complete, searchable history of every model version the system has ever produced. Any production model can be traced back to the exact data, code, and hyperparameters that generated it, and any previous version can be restored and redeployed from the registry.
Real-time inference via documented REST API
The FastAPI endpoint at localhost:8000/predict accepts JSON feature payloads and returns churn probability scores in real time, with auto-generated Swagger UI documentation making the API immediately consumable by downstream applications. Any tool or service that can make an HTTP POST request can integrate with the churn prediction system — no SDK, no custom client library required.
Tools
Tech Stack
| Technology | Purpose |
|---|---|
| scikit-learn | Logistic Regression model training and evaluation |
| FastAPI | REST API for real-time churn prediction inference |
| MLflow | Experiment tracking, model versioning, and artifact storage |
| EvidentlyAI | Data drift detection with automated retraining trigger |
| Docker / Docker Compose | Multi-container orchestration — API, monitoring, and MLflow services |
| GitHub Actions | CI/CD pipeline — automated testing and Docker image builds on every push to main |
| joblib | Model serialization and loading for serving |
| pandas | Data loading, imputation, splitting, and CSV I/O across all pipeline stages |
| pytest | Automated unit and integration testing within CI/CD pipeline |
| Python | Core language and system orchestration |