Active-Active Machine Learning Infrastructures For Training And Prediction
summary
Claims an active-active ML pipeline that does simultaneous training and prediction with streaming data and hot swaps models into production. Uses a Kafka message broker (Kinesis optional), a near‑real‑time data L, model repositories plus CI to hot‑plug new models and continuous retraining with sliding windows (minute/hour/day). Built running services that train in parallel and auto‑evaluate models; it works but struggles with data holes, concept drift and changing feature importance, requiring monitoring, window tuning and model management.
AI-written summary · about
chapters (13)
- 0:00Active-active ML infrastructures
- 2:59Concept drift in models
- 7:29Passive to active ML architectures
- 10:41Passive progressive ML explained
- 13:33Active pool ML workflow
- 16:30Active automated ML with streaming data
- 21:48Real-time training challenges
- 24:17Sliding window training
- 27:47Monitoring model performance
- 30:38Handling data holes and shifts
- 32:49Active-active infrastructure requirements
- 35:21Measuring AI performance
- 36:03Summary and challenges
related talks
25mContinuous ML Applications in Production
04mInterview
14mPerformance Optimization in High-Volume Transaction Systems
17mContinuous deployment for ML: The new software development lifecycle
30mReconceptualizing Machine Learning for the Real-time World
38mReal Time ML Pipelines in Multi-Tenant Environments
39mContinuous Delivery for Machine Learning — The Future of MLOps
31mRunning Machine Learning Analytics On Traces