Building an MLOps strategy from the ground up
summary
Practical guide to getting models into production: use vetiver (R/Python) to package pipelines with an input prototype, pin versions via pins, and deploy a REST endpoint built on FastAPI or a generated app.py plus Dockerfile to RStudio Connect, S3/GCS/Azure, or cloud container platforms. Vetiver adds monitoring (compute_metrics → MAE/MSE plots), partially automated model cards via Quarto, and handles scikit‑learn/xgboost/statsmodels/PyTorch (custom handler for Spark). Emphasis on reproducible, composable workflows and rollbacks via strong version hashes.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:59What is MLOps
- 7:55Data science workflow
- 11:03Versioning with vetiver and pins
- 17:12Deploying models via REST APIs
- 23:12Monitoring model performance
- 28:48Handling failures and retraining
- 32:43Ethical considerations
- 33:03Choosing MLOps tools
- 40:52Q&A
related talks
27mPractical MLOps for better models
19mDemystifying MLOps- 19mDemystifying MLOps
18mDemystifying MLOps with Vetiver
34mHolistic MLOps for better science
26mMLOps for the rest of us- A poor man's guide to putting models in production
25mBuilding Scalable MLOps Pipelines That Actually Work
31mMLOps In Practice – How To Run Your Machine Learning Models In Production At Enterprise Scale