Streaming Pipeline Anomaly Detection Benchmark: Kafka and Flink Telemetry Dataset

Benchmark dataset (v2) for ML-based anomaly detection on Apache Kafka and Apache Flink streaming pipeline telemetry. This version uses rotated cross-validation where each fold has a unique validation repetition, eliminating the validation bias present in v1. Changes from v1:- Cross-validation redesigned: validation repetition now rotates across all 5 folds- All 8 models retrained from scratch with corrected fold assignments- Feature category ablation expanded to 3 models (IF, OC-SVM, RF)- Derived feature ablation added (8 streaming-specific features have zero impact)- Censored detection latency for undetected episodes- Per-fold family average statistical test for RQ2 (avoids pseudoreplication) Contents:- Feature files at 4 window sizes (15s, 30s, 60s, 120s) with 475 features and rotated 5-fold CV assignments- Fault injection campaign manifests (80 runs across 8 fault types)- Model results for 8 models across 5 folds (F1, AUC-ROC, PR-AUC, precision, recall)- Per-fault detection performance (320 rows)- Detection latency with censored observations (640 rows)- Sensitivity analyses (prevalence, cooldown exclusion, window size)- Multi-model fusion results (IF, OC-SVM, RF with 4 strategies each)- Statistical test results (Friedman, Nemenyi, paired Wilcoxon, Cliff's delta)- All analysis scripts for full reproducibility Companion paper: "ML-based anomaly detection for real-time data streaming pipeline reliability" submitted to Future Generation Computer Systems (Elsevier).