Cost / Waste
Where the cluster's capacity goes, with an efficiency score, usage against requests, GPU use, nodes you could consolidate, spot risk, headroom forecasts and unusual usage.
Cost / Waste in the sidebar’s Tools accounts for the cluster’s capacity: what’s requested, what’s used, and what’s idle. It works in CPU and memory, not money, because Kubebay doesn’t know what you pay.
Efficiency score
One score, Good, Fair or Poor, from three parts:
- Provisioning (half the weight): how much of allocatable CPU and memory is requested. 80% or more earns full marks; the rest is burst headroom.
- Request coverage: how many containers set resource requests.
- Workload sizing: how many workloads are sized close to their p95 usage. It needs metrics-server or Prometheus. Without usage data it’s left out and says so, rather than guessed.
Capacity
- Idle capacity (cluster-wide): allocatable minus requests. It’s never charged to a namespace.
- System overhead (DaemonSets): what DaemonSets request, on its own line.
- Requested by namespace, down to each pod and container.
- By node: allocatable, requested, system overhead and idle for each node.
Workload usage vs requests
Each workload’s request, its observed p95 usage, how much is wasted, and the window it was measured over. It uses the same recommender as Right-sizing, but read-only. To change requests, use Right-sizing.
GPUs
Shown only on clusters with GPU nodes.
- GPU capacity: GPUs claimed against allocatable, per node.
- GPU utilisation: how busy each pod’s NVIDIA GPUs are, average and peak, and peak memory, from the DCGM exporter through your cluster’s Prometheus. A pod is marked underused, in use, too little data or no DCGM data.
Node consolidation
For each node: would its pods fit on the other nodes, by requests? The simulation honours nodeSelectors, required node affinity and NoSchedule and NoExecute taints. Each node gets a verdict:
- fits elsewhere: it could be drained;
- drain causes downtime: it runs a single-replica workload with no PodDisruptionBudget (see SPOF Radar);
- blocked: something stops a clean drain. A pod fits nowhere else, isn’t owned by a controller (so nothing would recreate it), has required pod affinity or anti-affinity or a host port (which the simulation doesn’t model), or a PodDisruptionBudget allows no disruption right now;
- skipped: the node isn’t a candidate: it’s not Ready, already cordoned, or a control-plane node.
The Why column names each blocker.
It’s read-only. Kubebay never cordons or drains a node from here.
Spot disruption risk
Workloads whose every running replica is on spot or preemptible nodes, so one reclaim wave can take them all:
- a single replica, which a PDB can’t protect, because reclaims don’t go through the Eviction API;
- several replicas with no PDB, so drains ahead of a reclaim aren’t coordinated.
History, forecasts and anomalies
These read Kubebay’s local usage history. A chip says how much history the cluster has, such as observed 09–18 local, weekdays only.
- Headroom forecast: the trend of the cluster’s daily peak against allocatable, and when it would run out. It needs at least 5 days, each with at least 4 well-sampled hours, and says how many more it needs until then. Missing hours are never filled in.
- Unusual usage: hours in the last day that stand out from the same hour on earlier days of the same kind (weekday or weekend). It needs at least 5 such days.
Because history is recorded only while Kubebay is open, the forecast is of the daily peak during the hours you had it open.