First Day Foresight: Anomaly Detection for Observability
Talkat KubeCon + CloudNativeCon Europe 2025 · on Wednesday, 2025-04-02· at 11:15· Level 1 | Hall Entrance N10 | Room E·
summary
Claim: useful anomaly detection runs on day one without full ML training by combining simple stats, synthetic data, and transfer learning. Uses z‑score rules in Prometheus/PromQL for immediate alerts, synthetic spikes and domain priors to bootstrap labels, an isolation forest demo served via KServe, and fine‑tuning of a Hugging Face LSTM with Kubeflow + KServe for scale. Example: a thumbnail‑latency incident would be caught earlier; local demo shows model fetch, serve, and alerting work end‑to‑end.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:26Reactive to proactive detection
- 2:50Stella Stash case study
- 5:17Day one anomaly detection
- 7:23Data instrumentation
- 8:48Feature engineering
- 9:30Cold start problem
- 12:19Demo: z-score in Prometheus
- 15:33Beyond statistical models
- 18:43CubeFlow infrastructure
- 20:33Scalable pipeline with CubeFlow
- 22:41Local setup with KServe
- 26:51Advantages and limitations
- 29:01Reducing failures lifecycle
- 30:47Conclusion and Q&A
related talks
38mWatch Everything, Watch Anything: Anomaly Detection
29mAlerting and Anomaly Detection – Best Friends Forever?
43mBeyond Logs: Real-Time Anomaly Detection with Machine Learning
26mProactive Failure Prediction Through Anomaly Detection
01h 21mBuilding a ML Model Observability pipeline
29mREAL-TIME ANOMALY DETECTION MADE EASY
30mPractical Anomaly Detection at Scale With PromQL
35mObservability’s Sixth Sense: Detecting Anomalies in Metrics