GWDG GPU Node Telemetry Dataset for Observability-Aware Early Warning of GPU Detachment Failures (2025-2026)

This dataset contains sanitized time-series telemetry from production GPU-equipped HPC nodes at the GWDG high-performance computing infrastructure, aligned with operator-curated GPU failure incidents recorded between January 2025 and February 2026. The dataset supports reproducible research on observability-aware early warning of GPU instability, particularly detachment-class failures where the dominant signal is not only numeric deviation but structural observability collapse such as metric disappearance, scrape degradation, and monitoring gaps. The telemetry is derived from Prometheus monitoring archives and includes multiple observability planes: • GPU telemetry collected via NVIDIA DCGM• Node and operating system telemetry from Prometheus node exporter• Monitoring pipeline telemetry such as scrape success and scrape duration• Scheduler state signals derived from Slurm exporter metrics Each incident window contains aligned pre-incident and post-incident telemetry intervals together with a curated incident catalog. Identifiers including node hostnames, GPU UUIDs, and exporter instances were pseudonymized using deterministic salted hashing. The dataset was sanitized to remove filesystem paths, environment references, URLs, IP addresses, and other sensitive textual patterns. The release contains no user data, job metadata, credentials, network topology information, or application-level traces. This dataset accompanies the study:"When GPUs Fail Quietly: Observability-Aware Early Warning Beyond Numeric Telemetry."