Real-Time ML services: From ideation to production
summary
Forecasting time series for a Real Time Dynamic Pricing service with 5–10 minute aggregated windows, a batch training job plus a streaming job consuming from Kafka, and features like rides requests available drivers zone demographics weather and events. Models train in TensorFlow (or XGBoost) with data versioning via DVC and pipelines on Kubeflow + Katib for automated hyperparameter search, CI and scheduled or performance-triggered retrains, plus Streamlit data apps and Grafana alerts — yielding reproducible deploys and faster delivery.
AI-written summary · about
chapters (14)
- 0:00Real-time ML services intro
- 3:15Ride-hailing demand supply
- 5:23Kathy's ML solution
- 10:21Traditional ML workflow challenges
- 12:10Why ML models fail in production
- 15:02ML engineering practices
- 16:51Reproducibility with data versioning
- 18:14Automated model retraining
- 19:59ML pipelines with Kubeflow
- 22:32Scaling hyperparameters with Katib
- 22:52Monitoring ML performance
- 24:18Stakeholder data apps
- 26:03Summary for real-time ML
- 27:05Q&A: Hyperparameters and testing
related talks
38mReal Time ML Pipelines in Multi-Tenant Environments
24mUnder Pressure: Applying ML in Real Time
38mSampling and Modelling the World in Real-Time for Mobility
22mScalable AutoML for Time Series Forecasting
25mContinuous ML Applications in Production
14mPerformance Optimization in High-Volume Transaction Systems
16mOperationalizing AI in Finance: Real-Time Analytics
34mTime series for scikit-learn people