Core Concepts
Dashboard: Overview · Capacity · Reliability · Cost
The per-cluster dashboard follows a scan-then-dive rhythm — a summary layer that answers "is anything on fire?", with deep instrumentation below.
On this page
Each cluster’s dashboard is organized into sub-tabs built around a scan → dive rhythm: a summary layer up top answers the urgent question at a glance, and the detailed panels below let you drill into anything. Open it with Dashboard in the top bar (its neighbour, Cluster Map, is the topology view). Overview and Capacity are always there; Reliability appears when the cluster ships Hubble flow data, and Cost (beta) when it ships OpenCost data. A tab with nothing to show is hidden rather than rendered empty.
| Sub-tab | Question it answers | Needs |
|---|---|---|
| Overview | Is everything fine right now? | Nothing beyond a connection; Metrics Server for live usage |
| Capacity | How is the cluster consuming, and is it sized right? | Metrics history from the agent for the card row and the trend charts |
| Reliability | What is the cluster actually serving? | Agent with Hubble enabled (Cilium); HTTP (L7) flows for the golden signals |
| Cost (beta) | What does it cost, and what is idle? | OpenCost data shipped by the agent; see Cost |
Every tab opens with the same header: a breadcrumb (Clusters / cluster / tab), the cluster’s name, and a line with the tab’s purpose, the cloud provider and region (detected from node metadata; they simply don’t render when they can’t be determined), the Kubernetes version and the node and namespace counts. On Overview, a line under the tabs lists which data sources are shipping samples for this cluster (the agent, Hubble, node-exporter and so on), and on a metrics-only cluster (an agent that ships metrics but gives KubeBolt no API access) a banner says the metrics dashboards work while the live resource views and operations are off, and how to turn them on.
The card row every tab opens with
Overview, Capacity, Reliability and Cost each open with a row of four cards built the same way: a large figure, one sentence saying what it means right now, a small chart chosen for that metric, and a mono caption with the breakdown and links to the lists behind it.
- A card lights up only when its number needs someone: a node down, pods not running, 5xx over the bar. The card glows from its left edge; the figure itself keeps its colour. A row where everything is fine has nothing lit, so a lit card is always worth reading, and a lit card names who (the node, the workload) with a link.
- A resource you can’t read keeps its slot. If your identity lacks access, the card says No access — insufficient permissions instead of showing a zero.
- The figures count up once. The first time you open a page after signing in (on cluster pages, once per cluster) the figures count up and the charts fill. Coming back, a data refresh, or reduced motion shows the final values.
Overview — is everything fine right now?
Overview has no range selector: everything on it is current state, and the trends live in Capacity. Its four cards:
| Card | Shows | Lights up |
|---|---|---|
| Cluster health | The score out of 100 on a half gauge, with a Healthy, Warning or Critical chip. The score adds up four checks (API server reachable, 25; nodes Ready, up to 25; Metrics Server available, 10; pods not failed or crash-looping, up to 40), then takes off 5 points per open critical insight (at most 25) and 2 per warning (at most 10). An open critical makes the chip Critical whatever the score. The card lists what took points off, and hovering shows every check | Never: the Insights card beside it carries the alert. Out of contact with the cluster it shows no score and a grey No contact chip |
| Nodes ready | Ready nodes over the total, and the three busiest (by the higher of CPU and memory) as double rings (CPU inside, memory outside) with “+N more”. Without node metrics, a Ready / Not ready list | Red when a node is not ready, amber when one is cordoned |
| Pods | Running pods over the active ones, and the pod count over the last 24 hours as a line (needs the agent; without it, a Running / Degraded list). The caption splits completed · degraded · not running, each linking to its filtered list | Red when pods are not running, amber when they are degraded (Pending, CrashLoopBackOff, partially ready) |
| Insights | Open insights as cases (one per rule and object; pods of one workload fold together), worst and oldest first, up to three rows; a rule with three or more cases becomes one row, rule ×N. The caption splits critical · warning · info | Red with a critical open, amber with a warning |
Completed Job pods are left out of the pods denominator, so a batch of finished Jobs never reads as missing pods. Muted insights are left out of the Insights card and the health score; a critical one still counts.
Below the cards:
- Resource efficiency — a capacity axis per resource with three stripes: used, requested-but-idle (striped), and free, with the limits as a tick. Each resource gets an efficiency score (used over requested), and a Right-size callout into Capacity when more than 30 % of the reservation sits idle. Live usage comes from Metrics Server, so it counts pods only.
- Recent Events, grouped, next to Workload Health. Each warning group has an Ask Kobi button (shown when Kobi is configured).
- Namespaces as tiles with their workloads and pod readiness: the first 12, those that need attention first, and Show all for the rest. The button says when hidden namespaces still need attention.
Capacity — a scan layer over the trends
The card row and the trend charts read metrics history, so they need the agent; without it the trends block shows a placeholder and the card row is not drawn.
- Card row: Peak CPU (its curve over the range, for the whole node: the OS and kubelet included), Peak memory (how much of capacity the peak took), Rightsizing opportunity (the reclaimable cores or GiB and the workloads that would hand back the most; with OpenCost connected it adds the ≈$/mo that is worth and links to Cost; until the cluster has the minimum usage history, 48 hours by default and set in Administration → System → General, it shows a dash and how much it has collected, see When recommendations appear), and OOMKills (when they happened, as marks on the range’s track). The peak cards light amber at 80 % of the cluster’s allocatable capacity and red at 90 %, naming the busiest node; OOMKills lights amber when there are any. The numbers come from the same series and queries the panels below plot.
- Cluster trends — CPU Usage, Memory Working Set, Network Activity and Filesystem Used, overlaid with deploy markers you can hide.
- Recent Deploys — Deployment rollouts in the range (each new ReplicaSet), with an Ask Kobi button per rollout. StatefulSet and DaemonSet rollouts are not listed.
- Top workloads by CPU, with ReplicaSet pods collapsed into their Deployment, next to Right-sizing Recommendations: per-workload reclaimable CPU and memory from the P95 of real usage. With cost data connected, the Cost tab prices those same recommendations.
- Recent OOMKills — containers whose last termination was an OOMKill, newest first, enriched with the container’s memory limit and restart count (you can tell a chronic crash loop from a one-off), with Ask Kobi.
Reliability — golden signals, honest colors
- Card row: Error rate 5xx (the traffic split by status class), Latency (this window against the previous one), Throughput (its curve over the range) and L4 drops (where they land). Comparisons are against the previous window of the same length, not the same time yesterday. The 5xx figure is drawn red from 0.1 % of requests, but the card lights only from 1 %; Latency lights amber when it rose more than 25 % and by at least 20 ms; L4 drops lights amber at a steady stream, one drop a minute or more on average over the range.
- Class-pinned colors: 5xx is always red and 4xx always amber, even when one class has no data in range.
- Panels below the cards: error rate split into 4xx and 5xx, top workloads
by traffic, error hot-spots (sorted by absolute error rate, so a small
service that always fails isn’t buried), top workloads by latency and
network drops (L4
droppedverdicts, the early warning for NetworkPolicy denials). - Latency is labeled as an average: the agent ships sum + count, not histogram buckets, and a tooltip says so instead of letting you assume a p99.
The Reliability tab is powered by Hubble flow data, so it appears only on
agent-connected clusters with Hubble enabled (hubble.enabled=true on the
agent chart). The golden signals need HTTP (L7) flows: when Hubble ships only
L3/L4 flows (Cilium running without its L7 proxy, as on GKE’s managed
Dataplane V2) the tab explains that instead of drawing empty panels, and
those flows still feed the Cluster Map’s Traffic layout.
See Connecting clusters.
Shared behavior
The selected time range (5 minutes to 30 days) is shared between Capacity, Reliability and Cost and persists while you drill into resources, but resets to the 15-minute default on a fresh browser session, so the daily scan never accidentally fires a 30-day query. The Monitor tab of each resource keeps a range of its own.
Cost (beta)
Spend, idle cost, cost per pod and money-priced rightsizing, from OpenCost data the agent ships. It is covered on its own page: Cost.