From Notebook Chaos to Production Gold
summary
A churn model that scores great in a notebook fails in production because training sees full, backfilled tables while serving sees partial, delayed rows and slightly different feature logic. The talk demonstrates two concrete failure modes: a payments feature that reads as $20 in training but 0 at inference due to late-arriving payment records, and a tickets feature where NaNs become zeros. The remedy is deploying Feast as a neutral feature layer over Snowflake and Redis using point-in-time joins for training and get_online_features for low-latency serving, plus scheduled materialization to keep stores consistent.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:27Notebook to production gap
- 3:14Two root causes explained
- 7:09Feature stores introduction
- 9:17Three feature store types
- 11:48Feature Store core concepts
- 15:20Point-in-time joins for training
- 18:55Online serving with materialization
- 20:22Molt’s full solution
- 21:47Summary and call to action
related talks
34mSolving the ML Pain You Forgot to Mention
29mMachine Learning in Production: What to do after you’ve trained the models?
29mWhen Worlds Collide: Productionalizing a Data Science Model
24mChallenges in building Machine Learning models in production
29mBuilding End to End Gen AI Powered Analytics: From Data to Production
59mWhy Great Models Fail: Lessons From 9 Years of Deploying ML Models
29mDemystifying Machine Learning in Production: Reasoning about a Large-Scale ML Platform
08mConsiderations for producing the best AI and Machine Learning models