Optimize Kubernetes Autoscaling Configuration

Kubernetes Autoscaler Configuration Expert helps configure and optimize Horizontal Pod Autoscaler (HPA), Vertical Pod Autoscaler (VPA), and Cluster Autoscaler. It treats pod replicas, container resource sizing, and node capacity as related parts of a production scaling strategy.

HPA examples scale deployments using CPU and memory utilization. The configuration includes minimum and maximum replica counts, separate scale-up and scale-down behavior, stabilization windows, percentage and pod-count policies, and policy selection. Custom-metric examples use external queue length and per-pod request rate to represent workload demand beyond resource utilization.

VPA examples define update behavior and resource boundaries per container. One configuration controls CPU and memory requests and limits while excluding a sidecar. Another uses recommendation-only mode for a PostgreSQL StatefulSet, with minimum and maximum resource recommendations rather than automatic changes.

Cluster Autoscaler configuration includes node bounds, delays after adding nodes, unneeded-node timing, cloud-provider configuration, node-group discovery, and a least-waste expansion strategy.

Use this skill when establishing autoscaling or investigating scaling that does not match observed demand. Resource requests provide the baseline for scaling decisions, and the guidance calls for multiple metrics, conservative ratios, and monitoring before changing thresholds.

Stabilization windows and appropriate delays address flapping. Gradual rollout, PodDisruptionBudgets, and alerts for scaling failures or resource exhaustion are part of the operational approach.

Workload patterns differ: web applications use CPU-based HPA, queue workers use queue depth, databases use VPA recommendations with manual tuning, and batch jobs use job-specific node pools. Microservices can combine HPA and VPA with explicit resource boundaries. Automatic vertical updates are therefore not the only pattern offered for every workload.

Inputs include the target Deployment or StatefulSet, resource requests and limits, replica bounds, metric targets, container policies, and node-pool settings. Outputs are YAML configuration examples and guidance for choosing scaling behavior.

A resource-request example pairs conservative CPU and memory baselines with burst limits and a readiness probe. The multi-tier strategy assigns immediate traffic response to HPA, longer-term resource optimization to VPA, and infrastructure capacity to Cluster Autoscaler.

Troubleshooting checks that the metrics server works and that target workloads declare resource requests. It also examines scaling events and compares actual pod resource usage with the configured baseline.

kubectl describe hpa

The configurations use Kubernetes resource metrics and custom metrics adapters. Monitoring examples include Prometheus ServiceMonitor configuration selecting kube-state-metrics, with a metrics endpoint and collection interval. The node-capacity example uses GKE-related cloud-provider and node-group discovery settings.