Realtime Resizing
Wave's automatic apply pipeline for Smart Sizing recommendations — In-Place Pod Resize (no restart, no rollout, no downtime) or Manifest Patch, per-container opt-in with VPA-style hysteresis.
In short
- • Every 10 minutes, Wave can apply Smart Sizing recommendations to your workloads automatically.
- • Three Apply Modes per workload: Off, In-Place (Kubernetes /resize, no restart), Manifest Patch (rollout-managed).
- • Auto-apply is opt-in at the container level. Nothing changes in your cluster until you turn it on.
- • VPA-style hysteresis (priority ≥ 10% OR CPU ≥ 50 mCPU OR memory ≥ 64 MiB) suppresses production noise.
Why better than VPA
Kubernetes' Vertical Pod Autoscaler applied recommendations by evicting and restarting pods. Its newer InPlaceOrRecreate mode (beta in K8s 1.33, GA 1.35) shares the /resize primitive Wave uses, but falls back to evict-and-recreate when the node lacks capacity — silently restarting your workloads.
Wave keeps the three Apply Modes as explicit user choices and never silently degrades In-Place to a restart. The recommendation pipeline upstream — per-container buffer / min / max controls, an OR-combined hysteresis that catches large-container drift VPA misses, and an Apply History panel — is the rest of the story.
Three Apply Modes
Off
Generate recommendations, never touch the workload. Use this to evaluate Smart Sizing before turning anything on.
In-Place Pod Resize
Calls the Kubernetes 1.27+ /resize subresource on running pods. No restart, no rollout, no downtime. Requires the InPlacePodVerticalScaling feature gate on K8s 1.27–1.32; enabled by default from 1.33, GA in 1.35.
Manifest Patch
Patches the workload spec. New pods pick up new resources on the next rollout. Safe everywhere — no feature gate required.
Supported on Deployment, StatefulSet, DaemonSet, Argo Rollout, and OpenShift DeploymentConfig.
Per-container buffer / min / max controls in Container Settings constrain the recommendation before the hysteresis check.
Before you turn auto-apply on
- • Hysteresis to suppress noise: applied only when priority ratio ≥ 10% OR CPU delta ≥ 50 mCPU OR memory delta ≥ 64 MiB (VPA-style thresholds, OR-combined).
- • In-Place is fire-and-forget per pod. If the kubelet rejects the resize because the node lacks resources, the pod is skipped and Wave retries on the next 10-minute cycle. Use Manifest mode if you need a guaranteed reschedule.
- • Manifest mode conflicts with Argo CD auto-sync — Argo will revert the patch. Turn auto-sync off for workloads where Wave should own resource values, the same way you already do for Autopilot.
- • On a Deployment, Smart Sizing auto-apply is automatically released when the workload transitions to Autopilot — Autopilot owns resource decisions for what it manages.
Close the loop on resource sizing
Recommendations are only half the value. Realtime Resizing is the automatic apply path that keeps your workloads right-sized every cycle — in-place, with no restart.
Read the operations guide