netdata

Why we included this project

The machine-learning layer is where netdata earns its keep: it trains lightweight unsupervised k-means models on every metric it collects, and rolling ensembles flag anomalies within about a second of them appearing. That approach skips deep-learning sophistication for what operators actually need, near-zero false positives through model consensus and no labeled training data, since per-metric behavior adapts automatically as workloads change. Anomaly bits live directly in the time-series database, so any chart can be queried for anomaly rates without separate ML infrastructure, and the tool correlates anomalies across a node's metrics, ranking them by severity to point at the likely root cause. Lean teams that want real-time monitoring plus practical anomaly detection without standing up a separate ML stack can run all of it as one self-contained system.