Important: Before You Begin

⚠️

Read this before installing. These behaviors trip up every first-time user.

Web Console Login

When the web console opens, it shows a login page.

Change these from the web console's user-management UI as soon as you're set up.

Smart Sizing analyzes workloads every 10 minutes.
On a fresh install, wait at least one analysis cycle before expecting recommendations to appear.

Read these before enabling Realtime Resizing. Full details in Realtime Resizing:

Manifest Patch conflicts with Argo CD auto-sync. Argo CD detects Wave's resource patches as drift and reverts them within minutes. Turn off auto-sync for any workload where Manifest Patch should own resource values (the same restriction as Autopilot, see Using ArgoCD under Wave Autoscale). In-Place mode is safe with Argo CD: it never changes the workload spec.
In-Place requires cluster support. Kubernetes < 1.27: not available, so use Manifest mode. 1.27 to 1.32: beta, requires the InPlacePodVerticalScaling feature gate. ≥ 1.33: GA, enabled by default. The console detects support at runtime and disables the In-Place option when the cluster can't do it.
In-Place is a one-shot call per pod, not a manifest change. Wave submits a single /resize request to each running pod. If the request is rejected (most commonly insufficient node resources), Wave skips that pod without evicting or rolling it out, and re-evaluates on the next 10-minute cycle.
Autopilot transition releases Realtime Resizing. When a Deployment is switched to Autopilot management, its Realtime Resizing is automatically turned off. Autopilot owns resource decisions from then on, so two writers never fight over the same fields.

All Wave Diagnosis tasks run on a 10-minute cycle:
- Cluster Resource Forecast: needs at least 7 days of metrics to produce a forecast.
- PV Capacity Forecast: needs at least 7 days of PV metrics to forecast volume exhaustion.
- Memory Leak Detection: needs at least 6 hours of metrics before it can analyze.
- CPU Utilization Analysis
- Pod Scheduling Delay Detection
- Idle Node Detection
On a fresh install, wait at least one cycle before expecting insights. Forecasts and leak detection take longer (see per-feature notes above).

Autopilot starts in fallback mode, an HPA-like CPU formula.
It switches to ML control after training completes.
Training requires CPU utilization data spanning 0-60%. Range matters, not duration; a workload that idles at 5% won't finish training until traffic exercises it up to ~60%.
The console shows a "Model Trained" badge per workload when ML is active.

Run new workloads in Monitoring Mode first.
The console renders a side-by-side chart: CPU-based (HPA-equivalent) vs ML-based decisions.
No scaling impact; it's pure observation.
Switch to Autopilot Mode when you're confident (see requirements below).

The target deployment must not have an HPA. Remove the HPA before switching to Autopilot Mode, since Wave takes over replica decisions and HPA would conflict.
Monitoring Mode is fine with HPA; this requirement only applies once you switch to Autopilot Mode.

ArgoCD-managed values (e.g., Deployment replicas) can conflict with values that Wave updates. Pick one of the following three methods:

Configure ArgoCD to ignore replicas changes. In Applications Detail → APP DETAILS → MANIFEST tab → edit, add ignoreDifferences for the spec.replicas field. See the ArgoCD ignoreDifferences documentation (opens in a new tab).
Comment out the replicas field in the Deployment manifest. Wave then manages replicas exclusively; ArgoCD has nothing to reconcile.
Configure via the ArgoCD ConfigMap (argocd-cm). Apply resource.customizations.ignoreDifferences rules cluster-wide so every Application ignores spec.replicas. Restart the ArgoCD pod for the change to take effect.