Lowering the barrier for ML monitoring
summary
Practical recipe: expose Prometheus metrics from a FastAPI app and visualize with Grafana, instrumenting Prometheus and Grafana plus Python client counters and histograms to capture traffic, errors, latency and a response distribution. Example pipeline uses scikit-learn transformers (onehot encoder, simple imputer) that emit per-category and missing-value counters so dashboards show surprising facts like request_count_total 8073 and features 9/10 with high missing counts. Demo runs from Docker Compose locally in a day and scales to Kubernetes in ~1–2 weeks, and the dashboard updates live.
AI-written summary · about
chapters (10)
- 0:00Introduction to ML Monitoring
- 1:07Importance and Purpose
- 2:27Monitoring Tools Overview
- 4:00Practical Scenario Setup
- 5:50Implementing Golden Signals
- 9:45ML Specific Monitoring
- 14:32Code Implementation
- 18:04Live Demo and Dashboard
- 24:43Understanding Model Decay
- 28:21Conclusion and Q&A
related talks
24mObservability in the MLOps Lifecycle with Prometheus
05mLightning Talk: Machine Learning Observability with Prometheus
32mDemocratizing Application Monitoring with Prometheus
24mPractical Kubernetes Monitoring with Prometheus
48mMetrics All the Way: Data Driven Operations
02h 43mPrometheus from A to Z
50mMetrics and alerting with Prometheus and Grafana
01h 12mDetecting Workload Anomalies with Prometheus and Machine Learning