Edge Fault Dataset

A dataset of QoS monitoring metrics collected during fault injection experiments on edge AI inference applications running on physical edge devices. The dataset is intended to support dependability and resilience research in edge computing environments and has been used for fault detection and classification tasks. Overview This dataset was created as part of a master's thesis at the Distributed Systems Group (DSG), Vienna University of Technology. It captures the behavior of two edge AI inference applications under 15 different fault types, enabling the study of how faults affect the Quality of Service (QoS) of edge applications. Format: CSV Size: 4.9 MB Rows: ~58,000 (≈ 16 hours of monitoring data at 1-second resolution) Fault injection experiments: 148 individual experiments across 16 fault types (including baseline) Target applications: 2 (YOLOv3 object detection, PocketSphinx speech recognition) Edge devices: 2 (ASRock V1000, Raspberry Pi 5) Dataset Structure The dataset is a single CSV file combining data from four experiment runs. Each row corresponds to a one-second Prometheus scrape interval. Columns / Features Feature Unit Description response_time seconds Time between sending the HTTP request and receiving the response processing_time seconds Duration of the inference function execution inside the app curl_exit_code — Exit code of the curl HTTP client http_response_status — HTTP response status code http_failed_requests requests/s Rate of failed HTTP requests http_successful_requests requests/s Rate of successful HTTP requests file_size MB File size of the HTTP request payload memory_usage GB Container memory usage cpu_usage CPU cores Container CPU usage network_receive MB Network data received by the container network_transmit MB Network data transmitted by the container fault_type — Label: the injected fault type (or no-fault for baseline) application — Target application (object-detection or speech-recognition) device — Edge device on which the experiment was run (asrock or rpi) Fault Types (Labels) Label Category Description no-fault Baseline Regular operation without fault injection pod-failure Container Makes a Pod continuously unavailable, simulating a container/Pod failure network-reorder Network Puts network packets into the wrong order network-loss Network Simulates packet loss by dropping a percentage of network packets network-duplicate Network Duplicates a percentage of network packets network-corrupt Network Corrupts a percentage of network packets network-partition Network Cuts off a Pod from the network, impeding all communication network-bandwidth Network Limits the network bandwidth for a Pod stress-cpu Resource Increases CPU load by executing CPU-intensive tasks inside a Pod stress-ram Resource Creates RAM pressure by allocating a large amount of memory to a Pod http-response-abort HTTP Aborts HTTP responses by forcibly closing the connection http-request-abort HTTP Aborts HTTP requests by forcibly closing the connection http-response-delay HTTP Injects a delay into HTTP responses http-request-delay HTTP Injects a delay into HTTP requests http-response-replace HTTP Replaces part of the content in HTTP responses http-request-replace HTTP Replaces part of the content in HTTP requests Experimental Setup Faults were injected using Chaos Mesh into two edge AI inference applications (YOLOv3 object detection and PocketSphinx speech recognition) running on a Kubernetes cluster on physical edge devices. QoS metrics were collected via Prometheus at a one-second scrape interval. For full details on the target applications, testbed hardware, software stack, and configuration, refer to the thesis. Citation If you use this dataset in your research, please cite the associated thesis: @mastersthesis{sepin2026evaluating, author = {Sepin, Stefan}, title = {Evaluating the impact of faults on the Quality of Service of edge-cloud applications}, school = {Technische Universit{\"a}t Wien}, year = {2026}, } The thesis is publicly available at: https://repositum.tuwien.at/handle/20.500.12708/227989 License This dataset is licensed under CC BY-SA 4.0.