Ship your Machine Learning Application #bbuzz
summary
A practical deployment pipeline is described that uses Apache Kafka topics named by preprocessor git commit and Kubernetes to promote models through build, test, canary and production stages; a prototype recommender showed a 14% increase in time on page but later served porn because test data mismatched live data. The system builds a canary instance and spot checker that joins live and canary predictions, routes disagreements to humans, stores vetted models in object storage (IBM/S3), and highlights limitations when preprocessors or rare labels change.
AI-written summary · about
chapters (11)
- 0:00Start
- 5:18Preprocessing Step
- 6:13Training Step
- 7:27Compare New And Old Model
- 8:09Publish New Model
- 9:04Service Build Step
- 10:38Production Overview
- 11:22Canary Deployment
- 12:17Spot Checking Select the correct document type
- 12:52Build up Test Data
- 15:47Limitations
related talks
40mHow to leverage the Apache Kafka Ecosystem to productionize Machine Learning
52mMachine Learning Data Pipelines
35mHow to deploy machine learning models to production (frequently and safely)
21mMastering the DevOps Machine Learning Pipeline
22mFrom experiment to production - a journey of a machine learning model
33mA Tale of Two Worlds: Canary-Testing for Both ML Models and Microservices
41mDeploying machine learning pipelines to production
58mThe path between developing and serving machine learning models.