Observability dashboards
Dashboards draw metrics and logs from the stores the seylo-observability
chart runs inside your cluster. Seylo reads them through the agent when a
page asks, and keeps none of the data; what Seylo stores is the dashboard’s
layout. Admin → Integrations → Observability stores shows whether each store is reachable
and, while one is not, the helm install command for the chart.
| Action | Role |
|---|---|
| View dashboards and templates | viewer |
| Create, edit or delete a dashboard; Use template; Build with AI | releaser |
Start from a template
Section titled “Start from a template”- Observability → Dashboards. Templates are listed below your own dashboards.
- Open one to preview it against your cluster.
- Use template copies it into a dashboard your organisation owns. Change it as you like; the template itself is unaffected.
Templates come in two lists, kept apart on the page. Each list folds: press its heading to close it once you know it, so your own dashboards keep the room. Whether a list is open or folded is remembered in your browser.
Templates that work with what the chart collects
Section titled “Templates that work with what the chart collects”These read only what the chart already collects — cAdvisor, the kubelet, kube-state-metrics, span metrics from traces, and container logs — so they show data as soon as the chart is installed.
| Template | What it shows |
|---|---|
| Kubernetes cluster overview | Capacity against requests and use, nodes, pods by phase, the busiest namespaces, restarts and failing workloads |
| Namespace workloads | Deployments, StatefulSets and DaemonSets in one namespace: replicas, CPU and memory against requests and limits, restarts, OOM kills, throttling, network and error logs |
| Pod and containers | One pod, container by container: CPU and memory against requests and limits, throttling, network, restarts, termination reasons and its logs |
| Persistent volumes | Space and inodes used, claims nearly full, requested size, storage classes and claims not bound |
| Autoscaling | HorizontalPodAutoscalers: current against desired, minimum and maximum replicas, those pinned at their maximum, and conditions that stop scaling |
| Jobs and CronJobs | Failures (also per CronJob), how long runs take, time since each CronJob last ran and last succeeded, and suspended schedules |
| Logs overview | Log and error volume for a namespace or one container, timeouts, connection failures, out-of-memory lines and the latest errors |
| Requests (rate, errors, duration) | Rate, errors and p50/p95/p99 duration per service and endpoint, from the span metrics derived from traces. Needs services that send OpenTelemetry traces |
| Cluster nodes | Node readiness, pressure, CPU and memory used and requested, and pod slots |
Kubernetes cluster overview and Cluster nodes read node series, which carry no namespace. Their node panels show data only to someone who may see every namespace.
Templates that need an exporter
Section titled “Templates that need an exporter”These read metrics the chart does not collect by itself. Each one says what to install or annotate, and until you do, its panels are empty.
From chart 0.2.19 the first two need nothing more: the chart runs node-exporter for Hosts, and finds ingress-nginx controllers by their labels for Ingress once the controller’s metrics are on. See host and ingress metrics. With an older chart, follow the table.
| Template | Install | Then annotate its pods with |
|---|---|---|
| Hosts (node-exporter) | Prometheus node-exporter as a DaemonSet | port 9100 |
| Ingress (ingress-nginx) | The ingress-nginx controller with controller.metrics.enabled=true |
port 10254 |
| JVM (Micrometer) | Micrometer’s Prometheus registry (Spring Boot: micrometer-registry-prometheus, and prometheus in management.endpoints.web.exposure.include) |
your management port, path /actuator/prometheus |
| Node.js (prom-client) | prom-client’s collectDefaultMetrics() with no prefix, served over HTTP |
your metrics port, path /metrics |
| PostgreSQL (postgres_exporter) | prometheus-community postgres_exporter beside the database, with a read-only monitoring role | port 9187 |
| Redis (redis_exporter) | oliver006/redis_exporter as a sidecar in each Redis pod, or beside it | port 9121 |
The annotations are:
metadata: annotations: prometheus.io/scrape: "true" prometheus.io/port: "9187" # the port from the table prometheus.io/path: "/metrics" # only where the table names a pathThe chart’s annotated-pods scrape job then collects them into the same store.
Each entry in this list tells you whether its data is there, for the cluster you are looking at: “Data found”, the metrics not seen in the last hour, or that none were seen. That check covers only namespaces you may see, so “not seen” can mean “not visible to you”. It never claims a metric is absent.
PostgreSQL connection strings stay with the exporter. Seylo never sees them.
Pickers
Section titled “Pickers”The pickers along the top of a dashboard (namespace, pod, service and so on) filter every panel. A picker can depend on one to its left: in Pod and containers, pick a namespace and the pod picker then offers only the pods of that namespace. When you change the namespace, the pod list refreshes.
Whatever a query says, you see only the namespaces you have access to. That is enforced by the store, not by the dashboard.
Network traffic is per pod
Section titled “Network traffic is per pod”A pod’s containers share one network, so traffic is reported for the pod, not per container. In Pod and containers CPU and memory are per container and network is per pod.
Pods on the host network report the node’s interfaces, not their own.
Usage metrics come from the same store
Section titled “Usage metrics come from the same store”With the chart installed, agents later than 0.2.4 read CPU, memory, network and throttling for Seylo’s own usage views and right-sizing from the chart’s metrics store, instead of scraping each node’s kubelet a second time. The numbers are the same; the kubelet is scraped once instead of twice.
This is also what gives a namespace-scoped agent usage metrics: it cannot read Nodes, but the store can. See choose an RBAC posture.
If the store cannot answer, the agent reads the kubelet itself for that round (where its permissions allow), says why, and tries the store again ten minutes later. Restart counts and OOM kills still come from the Kubernetes API, not from the store.
Build with AI
Section titled “Build with AI”Build with AI turns a description into a dashboard using the model your organisation chose under Nightshift → Settings: Seylo’s model or your own key (see Nightshift’s model). The model is told metric and log field names and the current layout — never your data. Its answer is checked like a pasted dashboard, metric names your store does not have are pointed out, and nothing is saved until you save it.
It runs in the background: the page shows how long it has been asking, with Cancel, and the answer says which model answered and how long it took. See the in-page assistants.
What the assistants answer
Section titled “What the assistants answer”Build with AI and Ask AI on Metrics and Logs answer about your clusters: their metrics, logs, traces, costs and workloads.
- A request that is plainly something else - an essay or a poem, a translation, homework, a maths problem, an unrelated coding exercise, a general-knowledge question - is refused before it reaches the model, and the message says what the assistant is for. A request that mentions your workloads, a signal or a metric name always goes through.
- The model is told the same, and answers anything else with Out of scope, without a second try.
- Each person may start Ask AI 100 times a day on each page, and Build with AI 40 times a day, counted over the last 24 hours. The query editors and the dashboard editor work without them. An admin can ask for more through Help → Contact support.