Anomaly Detection for Cloud Native Storage
summary
Implementation integrates volume metrics (IOPS, bandwidth, latency, capacity) from heterogeneous backends into a telemetry pipeline that exports to Prometheus server and forwards via Kafka to an ML module. The ML layer ingests ~6,000 volume-level points from LVM/Ceph and MongoDB tests, applies DBSCAN and Gaussian models to flag outliers tied to write bursts and gaps in scrapes, and posts alerts into Prometheus Alertmanager. A working demo triggers alerts on overwhelming write throughput; next steps are more data, cross-correlation with node/app metrics, and testing Random Forest/ARIMA.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:05What is anomaly detection
- 3:59Anomaly detection in storage
- 5:15Yahoo Japan’s storage infra
- 7:12Challenges in managing software-defined storage
- 8:44Roadmap to autonomous data platform
- 9:50HelmTree and Prometheus
- 13:10Metrics collection and exporters
- 15:37Prometheus alert manager
- 16:47Anomaly detection workflow
- 17:53Anomaly detection algorithms
- 20:26Testing Gaussian model with LBN
- 22:11Kubernetes cluster and MongoDB
- 25:33Live demo of HelmTree
- 30:11Future plans
- 31:23Q&A
related talks
31mAnomaly Detection on OpenStack Logs using M.L.
26mAnomaly Detection System for the CERN Cloud
30mBetter data quality with ML & visualization
43mBeyond Logs: Real-Time Anomaly Detection with Machine Learning
21mAnomaly Detection for Platform Logs using Embedded Context
34mML-Bound Anomaly Detection: Experiences Beyond Commodity Observability
40mAIOps: Anomaly detection with Prometheus
27mBeyond Logs: Real Time Anomaly Detection with Machine Learning