MLOps•2024
Iris MLOps Pipeline
Production-style CI/CD for an ML API with CML, Docker, Kubernetes, and MLflow.
Automated testing, container builds, deployment patterns, and model lifecycle thinking.
TL;DR
- End-to-end pipeline: train, evaluate, deploy.
- MLflow tracking for repeatable experiments.
- Containerized serving for consistent rollout.
Artifacts
Pipeline diagram
Train -> evaluate -> register -> deploy.
Model registry
Tracked runs and promotion to production.
Context
Manual ML deployments made it hard to reproduce results and ship updates quickly.
Problem
Inconsistent environments and ad-hoc releases slowed iteration and raised risk.
Approach
- Define pipeline stages with clear contracts.
- Track experiments in MLflow and register models.
- Package inference in Docker with pinned dependencies.
- Deploy via Kubernetes manifests.
Tradeoffs
- Invested in infra upfront to reduce future manual work.
- Chose Kubernetes for reproducibility over simpler VM deploys.
Testing and Reliability
- Unit tests for data and feature checks.
- Smoke tests for serving endpoints.
Deployment and Ops
- Automated rollouts via GitHub Actions.
- Rollback-ready deployment manifests.
Outcome
- Repeatable releases and cleaner handoffs.
- Faster iteration for model updates.
- Clear audit trail for experiments.
If I had two more weeks
- Add monitoring dashboards for drift and latency.
- Introduce canary deployments for new models.