Replication package: Measuring the Energy and CO2 Cost of ML-Based Anomaly Detection on Preprocessed OpenTelemetry Feature Tables

Data and code for the paper Measuring the Energy and CO2 Cost of ML-Based Anomaly Detection on Preprocessed OpenTelemetry Feature Tables (AICCSA 2026). Contents: the measurement harness; the three preprocessed OpenTelemetry feature tables (traces, metrics, logs); trained weights and fitted scalers for all fifteen model-signal configurations; and the measured results, covering per-inference energy and wall clock, detection quality, and the cascade, int8-quantization and region-shifting lever sweeps. Energy was measured on a bare-metal Intel Xeon Platinum 8275CL host, reading RAPL package and DRAM counters directly via /dev/cpu/N/msr with an idle baseline subtracted, so reported marginal energy isolates the workload. The feature tables derive from the chaos-engineered Kubernetes capture released with Evaluating ML-Based Anomaly Detection on Unified OpenTelemetry Telemetry, IEEE Access 14, 93576-93608 (doi:10.1109/ACCESS.2026.3705430), and are redistributed here so the package runs standalone.