Engineers guide for shepherding models in to production
summary
Checklist for production: define all model inputs and outputs and document ambiguous fields like location, plan for incomplete or dirty live data, and choose the integration point so the model has required inputs and runs early enough. Host models by splitting model code and model data, store artifacts in remote storage, expose a REST API with aliases for model versions for hot swaps and A/B tests, automate frequent retraining, monitor service health and prediction quality, and optimize for subsecond real‑time performance and cost through instance co‑location.
AI-written summary · about
chapters (9)
- 0:00Model production challenges
- 3:21Engineering’s role in deployment
- 7:35Automated model pipelines
- 12:14Model integration pitfalls
- 18:40Robust model inputs/outputs
- 19:02Hosting models for scale
- 25:00Dynamic updates and testing
- 28:34Automating retraining/monitoring
- 30:58Q&A
related talks
46mProduction Model Deployment
30mFrom Black Box to a Known Quantity: How to Build Predictable, Reliable ML-based
31mMLOps In Practice – How To Run Your Machine Learning Models In Production At Enterprise Scale
35mHow to deploy machine learning models to production (frequently and safely)
01h 05mMachine Learning Models and Productionization
42mImproving Machine Learning Development Reliability
30mWhen Good Models Go Bad: The damage caused by wayward models and how to prevent it
26mDataEngBytes 2023 - SYD-T2-03 - Akanksha Malik