KubeWatch: A Multi-Level Kubernetes Monitoring Dataset for Anomaly Detection

KubeWatch is a Kubernetes monitoring dataset for anomaly detection over multi-level runtime telemetry. It was collected on a dual cluster Kubernetes testbed in which a producer cluster runs monitored services and receives controlled anomaly injections, while a monitor cluster collects metrics, evaluates alerting rules, aggregates alerts, and records events. The dataset contains 11 days of monitoring data collected under continuous workload execution. The first 9 days capture normal runtime behavior, and the final 2 days contain controlled anomaly injections. KubeWatch covers three Kubernetes observation layers: node, pod, and control plane. Node and pod anomalies are injected under active workloads, while the control plane is retained as an observation layer for orchestration related responses and object state changes. This archive includes raw monitoring metrics, aggregated operational KPIs, injection records, alert histories, anomaly labels, and benchmark files. The KPI time series are derived from Prometheus recording rules and summarize raw metrics over 5 minute windows to reduce short term fluctuations and provide a stable operational view. Labels are constructed from controlled anomaly provenance and alert context, with refined anomaly windows aligned to the monitoring timeline. KubeWatch is intended to support empirical research on multivariate time series anomaly detection in Kubernetes environments. It provides standardized training data, test data, and test label files for evaluating anomaly detection methods across node, pod, and control plane telemetry.