Kubernetes testbed of "How Much Headroom Is Enough?": raw run data

Raw data of the Kubernetes testbed in Sect. 7 and Additional file 1, Sect. S11, of the article “How Much Headroom Is Enough? Overload Budgets for Predictive Autoscaling under Actuation Delay” (M. Kozhinov, M. Mazzara). Four 24-hour runs on a VK Cloud Managed Kubernetes cluster (21–26 October 2026) replayed 12 Huawei Cloud functions in real time under PAC-h, its standard-rank ablation, a tuned reactive controller, and the Kubernetes HPA. The archive also holds the capacity calibration, the pilot, every failed or superseded attempt, and the operator logs. The local engineering stage of 18 October 2026 is included as two separate archives. The code, frozen protocol, and analysis are in https://github.com/Maxkoz777/overload-budget-autoscaling, release v1.3.0 (https://doi.org/10.5281/zenodo.23701789). README.md describes the files and how to verify them; RESULTS_SHA256.txt lists the SHA-256 of all 428 files in the ZIP archive. Licence and attribution: CC BY 4.0. The replayed requests are derived from the Huawei Cloud 2023 serverless data release (https://github.com/sir-lab/data-release, CC BY 4.0; A. Joosen et al., ACM SoCC 2023, https://doi.org/10.1145/3620678.3624783). Changes: 12 functions selected from trace days 0–18, 147–165, and 168–184; per-minute request counts scaled so that one capacity unit equals one testbed pod (4 requests per second); counts replayed as individual requests at uniformly random times within each minute. The files contain no personal data and no credentials.