mini-mlops-platform (End-to-End, No Hype)

This repo demonstrates an end-to-end ML lifecycle: data versioning → training → model registry → containerized serving → monitoring → retraining triggers

No fancy model. The point is production thinking.

Stack

• PyTorch (tiny model)
• MLflow (tracking + registry)
• FastAPI (inference)
• Docker Compose (local platform)
• GitHub Actions (CI + scheduled retrain check)

Quickstart (local)

make venv
source .venv/bin/activate
make install

docker compose down -v
make serve
export MLFLOW_TRACKING_URI="http://localhost:5050"

make data
make train
make promote

Smoke test evidence (local)

Outputs saved under docs/smoke/:

• health.txt
• predict.txt
• metrics.txt
• docs.txt (Swagger UI HTML response)

Test inference

curl http://localhost:8000/health
curl -X POST http://localhost:8000/predict -H "Content-Type: application/json" -d '{"age":40,"income":90000}'
curl http://localhost:8000/metrics

Monitoring + retraining triggers

We keep baseline stats from training (mean/std per feature + baseline pred mean). A retrain can trigger when:

• enough requests exist (min_requests)
• feature mean shift (z-score) exceeds threshold (data drift proxy)
• prediction mean shift exceeds threshold (model behavior drift proxy)

Run:

python -m src.monitoring.retrain_decider
cat artifacts/retrain_decision.json

Latency vs throughput tradeoff

• Latency: single request time (FastAPI path)
• Throughput: requests/sec (batching, async, model loading strategy)
• This demo uses simple synchronous inference for clarity.