Machine Learning in Production: An Intro to MLOps.
summary
Claims that ~90% of ML models never reach production and that data, not code, creates most complexity. Describes concrete tooling and patterns: Kubeflow/MLflow pipelines for parameterized training experiments, serializing models to storage then serving via Seldon on Kubernetes (CRD, canary/AB/shadow rollouts) or framework-specific servers like TensorFlow Serving/TorchServe, plus Prometheus logging for monitoring and Elasticsearch for archived predictions. Explains detectors for outliers, distribution drift and adversarial inputs and how rollout metrics determine retrain or rollback decisions.
AI-written summary · about
chapters (12)
- 0:00Start
- 8:44Comparing Journeys
- 14:54ML is Different - Key Points
- 20:44Metrics Example
- 23:06Role of ML Ops
- 24:151 Training
- 27:05Training and CI
- 28:57Seldon ML Serving
- 29:552 Rollout and Monitoring
- 35:36Inference Graphs - Routing
- 36:39Advanced Monitoring - Post-Predict
- 43:16Advanced Topics Summary
related talks
13mMLOps: From AI Research to Production-Ready Systems
41mTHE STATE OF MLOPS - MACHINE LEARNING IN PRODUCTION AT ENTERPRISE SCALE
01h 05mMachine Learning Models and Productionization
14mGetting Machine Learning applications into production using MLOps and CD4ML
13mMLOps at Scale
25mBuilding Scalable MLOps Pipelines That Actually Work
26mAre you ready for MLOps? 🫵
28mMastering MLOps: Building Future Ready ML Systems