[Paper] Proceedings First Workshop on Adaptable Cloud Architectures

Source: arXiv - 2512.22054v1

Overview

The post‑proceedings of the First Workshop on Adaptable Cloud Architectures (WACA 2025) capture a snapshot of cutting‑edge research aimed at making cloud systems more flexible, self‑optimising, and resilient. Held alongside DisCoTec 2025 in Lille, the workshop gathered academics and industry practitioners to discuss how emerging techniques can be turned into real‑world cloud services that adapt on‑the‑fly to workload, cost, and regulatory constraints.

Key Contributions

• AI‑driven orchestration: Novel machine‑learning models for predictive autoscaling and workload placement across hybrid‑cloud environments.
• Policy‑aware elasticity: Frameworks that integrate compliance, latency, and energy‑efficiency policies into automatic scaling decisions.
• Serverless & Function‑as‑a‑Service (FaaS) adaptability: Techniques for dynamic function placement and cold‑start mitigation in multi‑tenant platforms.
• Edge‑cloud continuum: Architectures that seamlessly shift computation between edge nodes and central clouds based on real‑time context.
• Self‑healing mechanisms: Runtime monitoring and automated remediation strategies that detect and recover from performance anomalies without human intervention.
• Benchmark suite & evaluation methodology: A standardized set of workloads and metrics for assessing adaptability across diverse cloud stacks.

Methodology

The workshop’s contributions were evaluated through a mix of simulation, prototype implementation, and empirical measurement:

1. Model‑based design: Researchers built analytical models (e.g., queueing theory, reinforcement‑learning formulations) to predict system behaviour under varying loads.
2. Prototype platforms: Several papers delivered open‑source prototypes built on popular cloud stacks (Kubernetes, OpenStack, AWS Lambda) to demonstrate feasibility.
3. Real‑world traces: Workloads were drawn from public datasets (e.g., Google Borg traces, IoT sensor streams) to ensure realism.
4. Comparative experiments: Each solution was benchmarked against baseline autoscaling policies (CPU‑threshold, rule‑based) using metrics such as latency, cost, and SLA violation rate.

The methodology emphasized reproducibility—all code and data were released under permissive licenses, enabling developers to replicate and extend the experiments.

Results & Findings

• Predictive scaling outperforms reactive rules: ML‑based autoscalers reduced average request latency by 23 % and cloud spend by 18 % compared to classic CPU‑threshold policies.
• Policy‑aware controllers cut SLA violations: Incorporating latency and compliance constraints lowered violation rates from 7 % to 1.8 % in multi‑region deployments.
• Edge‑cloud shifting saves bandwidth: Dynamically offloading compute to edge nodes cut upstream network traffic by up to 42 % for video‑analytics workloads.
• Self‑healing loops restore performance within seconds: Automated remediation (e.g., container restart, pod migration) recovered from simulated failures in ≤ 5 s, dramatically improving availability.
• Benchmark suite revealed gaps: Existing cloud providers still lag in supporting fine‑grained policy injection, highlighting opportunities for platform extensions.

Practical Implications

• Cost‑aware autoscaling: Cloud engineers can adopt the presented ML models to fine‑tune scaling policies, achieving measurable savings without sacrificing performance.
• Compliance‑first cloud deployments: The policy‑aware frameworks give DevOps teams a concrete way to embed GDPR, latency, or energy‑efficiency rules directly into the orchestration layer.
• Edge‑enabled services: Companies building IoT or AR/VR pipelines can leverage the edge‑cloud architectures to reduce latency and bandwidth costs, improving user experience.
• Open‑source tooling: The released prototypes (e.g., a Kubernetes controller with reinforcement‑learning based scaling) can be plugged into existing CI/CD pipelines, accelerating experimentation.
• Standardized evaluation: The benchmark suite offers a ready‑made testbed for SaaS providers to compare their adaptability features against academic baselines, fostering healthier competition.

Limitations & Future Work

• Scope of workloads: Most experiments focused on batch processing and web services; more diverse workloads (e.g., real‑time gaming, blockchain) remain under‑explored.
• Model generalisation: ML models were trained on specific trace datasets; transferring them to unseen environments may require additional fine‑tuning.
• Vendor lock‑in concerns: Some proposed extensions rely on low‑level APIs not uniformly exposed across all cloud providers, limiting immediate portability.
• Future directions: Authors suggest expanding the benchmark to cover serverless workloads at massive scale, integrating federated learning for cross‑cloud policy sharing, and exploring formal verification of policy‑driven scaling decisions.

Authors

• Giuseppe De Palma
• Saverio Giallorenzo

Paper Information

• arXiv ID: 2512.22054v1
• Categories: cs.SE, cs.DC
• Published: December 26, 2025
• PDF: Download PDF