autoscaling

Autoscaling in cloud-native environments faces challenges including intricate microservice dependencies, highly dynamic and heterogeneous workloads, and poor cross-environment adaptability. This paper systematically surveys representative works published since 2020 and proposes a five-dimensional taxonomy—spanning infrastructure, architecture, scaling mechanisms, optimization objectives, and behavioral modeling—to enable fine-grained technical comparison and scenario-specific applicability analysis. We identify three emerging frontiers: large language model–driven autoscaling, microservice dependency–aware scheduling, and meta-learning–enhanced generalization—thereby bridging critical gaps in dynamic workload modeling and cross-platform adaptive scaling. By integrating performance profiling, workload feature extraction, anomaly detection, and dependency analysis, our work provides academia with a clear evolutionary roadmap and delivers to industry a practical, service-quality–aware technology selection framework that jointly optimizes resource efficiency and QoS guarantees.

Existing auto-scaling algorithms in edge computing suffer from poor SLA compliance, complex configuration, and delayed responsiveness. To address these challenges, this paper proposes a cloud-edge collaborative adaptive hybrid scaling mechanism. The method integrates proactive prediction with reactive feedback to enable microservice-granular, SLA-constrained dynamic resource orchestration, thereby ensuring low-latency and high-reliability service delivery. Innovatively, we design an SLA-aware elasticity policy engine that unifies a lightweight edge computing framework with hybrid cloud-edge resource coordination techniques. Experimental results demonstrate that the proposed approach significantly improves SLA attainment rate (+23.6%), reduces end-to-end latency variability (−41.2%), increases resource utilization by 18.3%, and cuts configuration parameters by 60%. Overall, it achieves an effective trade-off among performance, reliability, and operational complexity.

To address the challenges of SLA violation and delayed elasticity in cloud-based distributed services, this paper proposes FLAS—a lightweight elastic scaling architecture integrating proactive prediction and reactive response. FLAS makes three key contributions: (1) a high-order time-series trend forecasting model enabling accurate long-horizon prediction of workload and performance metrics; (2) a low-overhead, resource-utilization–based reactive mechanism for rapid failover under sudden load spikes; and (3) boundary-value–driven scaling decisions tailored to event-driven scenarios such as content publish-subscribe. Integrated into the E-SilboPS middleware, FLAS enables non-intrusive monitoring and fully automated scaling execution. Experimental evaluation demonstrates that FLAS consistently achieves ≥99% SLA compliance across diverse workloads, significantly improving elasticity responsiveness and resource efficiency under both normal and extreme conditions.

This work addresses the limitations of existing Kubernetes autoscaling mechanisms—such as Horizontal Pod Autoscaler (HPA) and KEDA—which rely on reactive control and struggle to meet latency SLOs for Node.js applications due to lagging CPU or event-loop metrics and feedback interference from scaling actions. To overcome this, the authors propose a predictive autoscaling approach that leverages cluster-level aggregated load metrics to eliminate feedback noise and employs short-term time-series extrapolation to forecast workload at the moment new instances become ready, enabling proactive, future-oriented scaling. The method introduces an invariance-based metric model comprising three functions and a five-stage data processing pipeline, achieving near-target instance utilization under both steady-state and bursty traffic conditions. Evaluated empirically, it attains a median steady-state latency of only 26 ms, substantially outperforming KEDA (154 ms) and HPA (522 ms).

To address response latency, inaccurate prediction, and complex configuration in auto-scaling microservices under stringent SLAs (e.g., low latency, high reliability) in edge computing, this paper proposes a hybrid elasticity mechanism integrating time-series forecasting with real-time feedback. We innovatively design a dual-mode framework that synergistically combines reactive and proactive scaling, and— for the first time—deeply embed a lightweight machine learning–based demand prediction model into the native Kubernetes control loop, enabling dynamic threshold adjustment and SLA-driven adaptive scaling. Evaluation on a real-world edge testbed demonstrates that our approach reduces SLA violation rate to 6%, a >70% improvement over state-of-the-art methods; it also significantly enhances end-to-end low-latency guarantees and system availability for multiple IoT applications.

Traditional Horizontal Pod Autoscalers (HPAs) struggle to handle resource disturbances caused by failures, cyberattacks, or operational activities, often leading to service unavailability, resource wastage, and control instability. This paper proposes SecureSmart HPA—a disturbance-aware horizontal autoscaling mechanism tailored for microservice architectures. Its core contributions are threefold: (1) a novel dynamic scaling decision framework jointly driven by real-time disturbance detection and quantitative resource-wastage assessment; (2) a cross-service shared-resource scheduling strategy to enhance elasticity and utilization under resource constraints; and (3) adaptive control leveraging monitoring feedback, dynamically adjusted thresholds, and lightweight disturbance modeling. Experimental evaluation under 25%–75% disturbance intensity demonstrates that SecureSmart HPA reduces CPU overload by 57.2% and improves resource allocation efficiency by 51.1% compared to Smart HPA, while significantly enhancing system stability and response efficiency.

Kubernetes’ native autoscaling mechanisms—relying on reactive decision-making, underutilizing application-layer signals, and employing opaque control logic—frequently violate SLOs and waste resources. To address these limitations, we propose an AIOps-driven, multi-signal collaborative autoscaling framework that jointly optimizes for SLO compliance, cost constraints, and lightweight time-series demand forecasting. This work establishes the first SLO-first, cost-aware, and safety-guaranteed unified control paradigm with inherent interpretability. By integrating multidimensional metrics into a unified model and embedding a closed-loop feedback controller, our approach ensures transparent scaling decisions and auditable operational traces. Experimental evaluation demonstrates a 31% reduction in SLO violation duration, a 24% acceleration in scaling responsiveness, and an 18% decrease in infrastructure costs—all while preserving system stability and full operational traceability.

This work addresses the inefficiencies in cloud-native autoscaling caused by misalignment between business policies and resource scheduling, as well as the lack of coordination between pod and node scaling, which often leads to resource waste and performance degradation. To resolve these issues, the authors propose MAS-H2, a three-tier hierarchical multi-agent system that introduces, for the first time, a hierarchical multi-agent architecture to cloud-native autoscaling. MAS-H2 integrates strategic, planning, and execution agents to formalize business objectives, enable joint proactive scaling of pods and nodes, and achieve zero-downtime strategy migration. Implemented as a Kubernetes Operator, the system combines time-series forecasting, global utility optimization, and multi-agent control. Experimental results on GKE demonstrate that CPU utilization remains below 40%, sustained load is reduced by over 50% compared to HPA, peak CPU load during traffic bursts drops by 55%, and zero-downtime infrastructure migration is successfully accomplished.

This work proposes a proactive autoscaling framework for Kubernetes based on deep reinforcement learning, addressing the limitations of conventional reactive approaches that often lead to resource wastage or performance bottlenecks. The proposed method uniquely integrates Deep Q-Networks (DQN) with Long Short-Term Memory (LSTM) networks: the LSTM component forecasts future workload demands, while the DQN leverages these predictions to generate forward-looking scaling decisions. This framework is seamlessly integrated into the Kubernetes platform, enabling intelligent and efficient elasticity in cloud-native environments. Experimental results demonstrate that, compared to existing solutions such as Horizontal Pod Autoscaler (HPA) and KEDA, the approach significantly reduces resource consumption while maintaining stringent performance requirements.

This study addresses how cloud infrastructure failures distort performance metrics, thereby misleading autoscaling systems and leading to resource misallocation, increased costs, or degraded service reliability. Through controlled simulations, the authors systematically evaluate the impact of four common failure types—including storage and network routing issues—on both vertical and horizontal scaling strategies across varying instance configurations and SLO thresholds. The work presents the first quantitative analysis of how such failures bias scaling decisions, revealing that horizontal scaling is particularly sensitive to transient anomalies. It further proposes design principles to distinguish genuine workload changes from failure-induced artifacts. Experimental results demonstrate that storage failures can incur up to $258 in additional monthly costs under horizontal scaling, while routing anomalies consistently cause resource under-provisioning, offering empirical foundations for building fault-tolerant autoscaling mechanisms.

This work addresses the inefficiency of manual performance tuning in cloud-native stream processing systems, which heavily relies on expert experience. To automate and accelerate configuration optimization, the authors propose an experiment-driven approach that integrates Latin hypercube sampling, simulated annealing, and hill climbing into a three-stage search strategy. This method is deeply coupled with the Theodolite benchmarking framework to automatically orchestrate experiments on Kubernetes and preemptively terminate underperforming configurations. Evaluated on Kafka Streams, the approach efficiently explores the configuration space and identifies settings that substantially outperform default configurations, achieving up to a 23% improvement in throughput. The study demonstrates a practical and effective pathway toward automated, high-efficiency tuning of stream processing systems in cloud-native environments.