Experimental Telemetry Dataset and Codebase for Autonomous Self-Healing in Kubernetes Clusters Using a Hybrid Resilience and Q-Learning Rollback Model

This repository contains the complete raw physical testbed telemetry datasets, Kubernetes YAML manifests, Python analysis pipelines, and deployment automation scripts from our experimental campaign evaluating the "Hybrid Resilience (H/R)" Testing Model in Kubernetes environments. The codebase and data validate an integrated, multi-layered deployment framework featuring AI-driven pre-deployment risk estimation, controlled chaos validation in staging, and Istio outlier-detection self-healing coupled with a Q-learning MDP adaptive rollback controller to minimize application downtime and MTTR during severe deployment and cluster failures. 📄 ASSOCIATED MANUSCRIPT This dataset and codebase support the findings presented in the research article: Haranath Rakshit and Subhasis Banerjee, "Autonomous Self-Healing in Kubernetes Clusters Using a Hybrid Resilience and Q-Learning Rollback Model," 2026. 📂 REPOSITORY STRUCTURE HR_Experiments/ ├── 00_Workflow/ (Master replication guide & indexes) ├── 01_Foundation/ (Cluster setup: Helm, Istio, Litmus) ├── 02_AI_Prediction/ (Phase 1: ML dataset, FastAPI service) ├── 03_Chaos_Engineering/ (Phase 2: Chaos manifests & RS calculator) ├── 04_Istio_SelfHealing/ (Phase 3: Outlier detection & VS rules) ├── 05_MDP_QLearning/ (Phase 4: MDP Q-learning engine) ├── 06_Results/ (Phase 5: Aggregate results & ROI analysis) ├── 07_Revalidation/ (Environment A: Physical x86_64 metrics) ├── 08_RaspberryPi_Replication/ (Environment B: ARM64 edge logs & scripts) └── 09_Cross_Environment_Analysis/ (Cross-architecture comparison tables) 📊 KEY TELEMETRY COLUMN DEFINITIONS cpu_util / memory_util: Node CPU and memory utilization metrics scraped from the Kubernetes cluster API (measured in percentage or millicores/bytes). latency_ms: Client-side round-trip response latency for application transactions measured in milliseconds (ms). error_rate: Percentage of HTTP 5xx errors returned by application pods during a Prometheus scraping window. failure_probability_FP / FP_pred: Machine-learning predicted deployment failure probability score (ranging from 0.0 to 1.0). resilience_score_RS: Quantitative stability score calculated based on normalized MTBF and MTTR metrics. chaos_impact_D: Quantitative metric measuring performance degradation under active chaos fault injection. 🛠 USAGE & REPRODUCIBILITY To audit the datasets, evaluate the pre-deployment failure predictions, and automatically reproduce the classification metrics (e.g., F1-score of 0.8717), self-healing recovery benchmarks (e.g., 66.7% downtime reduction, 60% MTTR reduction), and the figures found in the associated manuscript, navigate to the phase directories and run the python scripts: To run the multi-classifier pre-deployment failure prediction benchmark: cd 02_AI_Prediction/ python 05_multi_model_benchmark.py To run staging fault injection and calculate the Resilience Score: cd 03_Chaos_Engineering/ python 05_calculate_rs.py To execute the MDP Q-learning training iterations: cd 05_MDP_QLearning/ python 01_mdp_engine.py To generate the baseline comparative downtime figures: cd 06_Results/ python 03_baseline_comparison.py Environment: Python 3.10+ (with packages: Scikit-Learn, Numpy, Pandas, FastAPI, Uvicorn) and Kubernetes cluster v1.31+ with Istio service mesh v1.22.0 and LitmusChaos v3.x.