Production Model Deployment
summary
Describes concrete deployment choices: put models behind an API scoring service (stateless, scale by load balancer), write batch outputs to files for cron-driven jobs, or publish parameters as a package; recommends serialization options like scikit-learn pickle or PMML with OpenScoring. Advises monitoring via AB tests, feature and prediction distributions, and rebuilding pipelines on a cadence (often daily retrain). Solves latency with a feature store or materialized predictions, warns against slow engineer handoffs and rewriting models in C.
AI-written summary · about
chapters (38)
- 0:00Intro
- 0:34Juliets background
- 5:12Agenda
- 6:26Lifecycle
- 6:43Black boxes
- 7:04Featureization
- 8:11Deployment as Sharing
- 9:56Models
- 12:39Model App
- 12:56Problem Serialization
- 14:20Does it Function
- 15:11Does it Work
- 16:00AB Testing
- 16:48AB Testing Forever
- 18:55Deployment
- 19:53Deployment schedule
- 21:17When to deploy
- 21:38Machine learning pipelines
- 22:21Lambda architecture
- 23:29Engineering requirements
- 25:31Model throughput
- 27:59Feature store
- 29:44Model output
- 30:43Evergreen solution
- 31:51The handoff
- 32:37Conways Law
- 33:12Data Scientist vs Software Engineer
- 33:32Data Science Departments
- 34:06Machine Learning
- 35:01Communication
- 35:14Model Handoff
- 36:13Clean Interfaces
- 36:39Serialization
- 37:39PMML
- 39:03Limitations of PMML
- 40:49General Questions
- 41:46Team Structure
- 42:29QA
related talks
48mDeploying Machine Learning Models in Practice
42mDeploying ML models and all the things that go wrong
31mEngineers guide for shepherding models in to production
39mApplying deployment oriented mindset for building Machine Learning models
31mMLOps In Practice – How To Run Your Machine Learning Models In Production At Enterprise Scale
29mHow to automate the deployment of ...
51mSpeeding up deployment of machine learning models
34mOpen Standards For Machine Learning Model Deployment