Wave - Amazon EKS Operations: Scaling, Sizing, Karpenter, Diagnosis

WHAT AMAZON EKS LEAVES TO YOUR TEAM Amazon EKS runs your clusters and supports HPA and Karpenter. Three problems are still solved by hand:

• Failures: a container restarts and its last logs are already gone. Your team cannot tell a memory leak from normal load.
• Cost: there is no basis for Request/Limit, so teams over-allocate to be safe, and nobody lowers the values later.
• Tuning: every workload needs its own HPA thresholds, the values go out of date when traffic changes, and event peaks are still scaled by hand.

Wave (formerly Wave Autoscale) solves these with data. It collects metrics, logs, and Kubernetes state together and supports the Amazon EKS operations lifecycle in one console: Monitor, Diagnose, Optimize, Control, and Report & Alert.

SAFE TO INSTALL

• Runs only inside your EKS cluster (Wave Core and a Metrics Agent at 100m CPU / 50Mi per node). No external egress, so it also installs in air-gapped networks.
• Apply Mode is Off after install: Wave only recommends until you turn it on.
• No dependency on metrics-server or Prometheus. Wave analyzes every 10 minutes and gives the first recommendation within 24 to 48 hours.
• Every recommendation shows its calculation. Original values are saved, so you can restore them in one step.
• Validated on clusters of up to 1,000 nodes.

WAVE SIZING: Request/Limit from data

• Smart Sizing: per-container CPU and memory recommendations from 7+ days of actual usage. Flags under-provisioned and over-provisioned containers.
• Realtime Resizing: applies the 10-minute recommendation to opt-in workloads. Uses in-place resize without a restart on Amazon EKS 1.27+, or a manifest patch on older clusters.
• Smart Sizing Report: current and recommended resources, stability, and potential savings in one PDF.

WAVE DIAGNOSIS: answers the operator's questions

• Container Failure Diagnosis: keeps the exit reason and last logs before a restart, Pod deletion, or log rotation removes them.
• Memory Leak Detection: finds steady memory growth and warns of OOM risk before it happens.
• Pod Scheduling Delay Detection, CPU Utilization Analysis, Idle Node Detection, and Unused PV Detection.
• Cluster Resource Forecast and PV Capacity Forecast: warn 7 to 60 days before CPU, memory, pod, or storage limits are reached.

WAVE AUTOSCALE: scaling on workload performance

• Autopilot: ML-based horizontal Pod scaling on a model of each workload's performance. Decisions within 10 seconds, about 2x faster than HPA (about 30 seconds), with no threshold tuning.
• Autopilot Scheduler: cron and calendar schedules for known events such as campaigns and business hours.
• Min/Max Recommendation: replica bounds from real demand.

WAVE KARPENTER: Karpenter operations (Amazon EKS and OKE)

• Karpenter Dashboard: nodes, hourly cost, Spot ratio, and wasted cost in one view.
• Node Warmup: prepares nodes before Pods go Pending, so Pods do not wait 30 to 90 seconds for a new node.
• Spot Workload Placement: Prefer Spot or Require Spot per workload. Up to 70% lower node cost.

WAVE FLOW: traffic protection at peak

• Priority-based traffic protection for Istio Sidecar, Istio Ambient, and Envoy Standalone. The lowest of four priority levels is throttled first, so critical traffic such as payments stays served. Does not replace Istio.
• NetFUNNEL integration: a virtual waiting room driven by workload state.

PLATFORM

• One console for many Amazon EKS clusters. GitOps through CRD mode with Argo CD.
• PV auto-expansion and policy-based PV cleanup.
• Slack and webhook alerts, Prometheus and Grafana integration, a full REST API, and OIDC single sign-on.

USE CASES

• BC Card, financial services (13 million members, 100+ workloads, private cloud Kubernetes): replaced HPA tuning with Autopilot, Autopilot Scheduler, Min/Max, and Smart Sizing. It handles the same peak traffic on 60% of its previous resources, and event-peak response time dropped from about 10 seconds to about 1.5 seconds.
• SaaS company on Amazon EKS (Korea, Japan, APAC): customer events set the traffic peaks. Autopilot Scheduler switches to an event preset minutes before each event, and Autopilot still reacts within 10 seconds.
• Global service that must keep HPA on: HPA keeps deciding the replica count, and Smart Sizing sets container requests, with Aggressive Recommendations to cut cost.

PRICING AND FREE PILOT

• Free for clusters under 200 vCPU (the sum of allocatable vCPUs across all worker nodes). Above 200 vCPU, usage is metered through AWS Marketplace.
• Free pilot: Wave engineers install Wave on your EKS cluster at no cost and, in 1 to 4 weeks, deliver four PDF reports from your own data: Smart Sizing, Wave Diagnosis, Autoscaling, and Cluster Resource.

Installs on Amazon EKS 1.26+ with a Helm chart. Does not replace your CNI, control plane, or service mesh. AWS EKS Service Ready Partner.