AgileSphere: Predictive-Reactive Kubernetes Cluster Autoscaling With Warm-Pool Node Promotion
Container orchestration platforms have made cloud-native deployment routine, but autoscaling at the cluster level remains a persistent source of capacity-pressure service-level objective (SLO) violations under bursty workloads. Existing autoscalers either react slowly to capacity exhaustion, paying cloud cold-start latency on each scale-up, or scale only at the container level, leaving pending pods on saturated nodes. We present AgileSphere, a predictive-reactive cluster autoscaling framework that combines (i) ScalePlanner, an iTransformer-based capacity forecaster trained on the Alibaba 2022 cluster trace; (ii) ScaleBit, a magnitude-aware reactive controller that promotes pre-cordoned reserve nodes to active capacity in seconds—a warm-cache <inline-formula> <tex-math notation="LaTeX">$p_{95}$ </tex-math></inline-formula> of 4.30 s, rising to ~14 s only on a first cold image pull (vs. 60–180 s for cloud provisioning); and (iii) AgileServer, a monotone-max reconciler that arbitrates between the two desired-state sources. We implement the prototype on a live Kubernetes testbed and evaluate it under five conditions, including a Kubernetes Cluster Autoscaler (KCA)-equivalent port with modeled cloud cold-start latency. On a sustained-pressure workload across <inline-formula> <tex-math notation="LaTeX">$N=10$ </tex-math></inline-formula> seeds, AgileSphere reduces mean pending-pod-seconds (PPS) by <inline-formula> <tex-math notation="LaTeX">$23.3\times $ </tex-math></inline-formula> versus vanilla Kubernetes and <inline-formula> <tex-math notation="LaTeX">$2.2\times $ </tex-math></inline-formula> versus KCA at near-identical active node-seconds. A static-10-node over-provisioning ablation rules out the trivial alternative: AgileSphere delivers 80% lower backlog at comparable active node-seconds (within 5%) to a pre-uncordoned 10-node cluster, so the gain is not from “more capacity” but from “capacity at the right time.” Beyond mean improvement, the predictive arm contributes a <inline-formula> <tex-math notation="LaTeX">$5.2\times $ </tex-math></inline-formula> tighter seed-to-seed standard deviation than pure-reactive scaling, identifying a stability advantage that prior cluster-autoscaler work has not characterized. On an exploratory transient-pressure bursty workload (<inline-formula> <tex-math notation="LaTeX">$N=3$ </tex-math></inline-formula> seeds), AgileSphere remains statistically indistinguishable from the best reactive-only baseline on mean PPS (within 17%) while improving over KCA on both backlog (<inline-formula> <tex-math notation="LaTeX">$8.6\times $ </tex-math></inline-formula> lower) and active node-seconds (4% fewer) on the same workload. The reserve-pool promotion primitive achieves a warm-cache <inline-formula> <tex-math notation="LaTeX">$p_{95}$ </tex-math></inline-formula> of 4.30 s, approximately <inline-formula> <tex-math notation="LaTeX">$20\times $ </tex-math></inline-formula> tighter than modeled cloud cold-start latency; we also quantify the infrastructure cost of holding such a warm pool, which raises total provisioned capacity even where active node-seconds are comparable. These results show that hybrid predictive-reactive autoscaling delivers lower sustained-pressure backlog and higher operational stability than single-paradigm baselines, while remaining competitive with the best reactive-only configuration on transient workloads.