Skip to content

Namespace-scoped installs

An agent is namespace-scoped when it was installed with a list of namespaces. That is rbac.scope=namespaced with watchNamespaces, or a fixed watchNamespaces list on a cluster-wide install (see choose an RBAC posture). The same rules apply to a person whose access is limited to some namespaces, and to every deploy token, which is always limited to one.

The chart is cluster-wide, and the reads are not

Section titled “The chart is cluster-wide, and the reads are not”

The seylo-observability chart collects from the whole cluster. VictoriaMetrics scrapes every node, vlagent reads every pod’s logs, and kube-state-metrics reports every namespace. That is why the chart needs its own read-only ClusterRoles, and why the cluster owner, not the namespace owner, installs it.

What Seylo reads back is narrowed in two places:

  1. The agent intersects whatever Seylo asks for with its own namespace list. Seylo cannot widen it.
  2. The store applies the resulting namespace list as a filter on every query. That filter is added by Seylo, never taken from the query text, and a request that tries to set it is refused.

The agent’s own Role does not change. Installing the chart gives the agent no new Kubernetes permission. The stores’ scraping uses the chart’s ClusterRoles, not the agent’s.

Seylo proves the filter before it relies on it

Section titled “Seylo proves the filter before it relies on it”

A store installed by someone else might be an older version. It might sit behind a proxy that drops arguments, or not be a Victoria store at all. In each case it could silently ignore the namespace filter, so Seylo tests the filter instead of assuming it works.

For each signal, Seylo asks the store the same count twice. The first time is unfiltered. The second is filtered to a namespace nobody has, seylo-enforcement-probe-0000. The filter is enforced if the filtered count is zero and the unfiltered one is not. The metrics probe counts namespaced series. The logs and traces probes look back one day, so a quiet namespace does not count as “no data”.

Admin → Integrations → Observability stores shows the verdict in the Namespace filter column:

Verdict Who is affected
namespace filter enforced Nobody. Scoped agents, limited people and tokens can query
namespace filter NOT enforced Refused. The store returned data for a namespace that does not exist
no data yet to test the filter Refused until the store holds data to compare
filter check failed Refused, with the failure as the reason

The verdict is kept for ten minutes. Check again runs it now.

Through a namespace-scoped agent, the check applies to everyone, including people who can see every namespace. The agent applies its scope to every query, so a store that ignored the filter would show those people namespaces that the platform team kept out of the agent’s reach.

People who can see every namespace, on a cluster-wide agent, are never affected. There is nothing to hide from them.

Cluster-wide agent Namespace-scoped agent
Metrics, logs, traces, APM for the agent’s namespaces ✅ ✅ once the filter is proved for that signal
Node series: node CPU and memory, readiness, pressure, allocatable ✅ for people who see every namespace ❌ These series carry no namespace, so no filter can admit them
Kubernetes cluster overview and Cluster nodes dashboards ✅ in full Workload panels only. Node panels are empty
Usage, right-sizing, network and throttling, with the chart and agent 0.2.5 or later ✅ ✅ read from the store, for the agent’s namespaces
Usage without the chart (default kubelet source) ✅ the agent scrapes the kubelet ❌ unknown. A Role cannot grant access to Nodes
Managed vCPU Measured for the namespaces Seylo manages Measured for the agent’s namespaces only
Spans that the collector could not match to a pod ✅ for people who see every namespace ❌ no namespace, so never shown
Pod-failure alerts ✅ ✅ They come from pod status and events, not the store

A trace that crosses namespaces shows only your part of it. The Traces page says when spans were left out.

For the alternatives to the chart on a namespace-scoped agent (a Prometheus you already run, or metrics-server), see the one real consequence of namespace scoping.

When Seylo cannot measure something, it shows the figure as missing and says why. It never shows a zero. “This container used no network” and “this source cannot see network” are different facts, and a right-sizing decision made on the wrong one costs you money or an outage.

On a namespace-scoped install, this is how it shows:

  • Usage shows n/a for a figure the source did not report, and the Data source card names the reason. Without the chart, an agent that cannot list Nodes reports that the kubelet source cannot be scraped, and why.
  • A store answer with no CPU series while containers are running is treated as no data, not as zero use. The agent falls back to its own sources and says so. See how usage is collected.
  • Managed vCPU is not measured when the store cannot be read, or has no kube-state-metrics series. It is never estimated, and an unread day is left out of the month rather than counted as zero.
  • Dashboard templates that need an exporter report metrics as “not seen”, never “absent”. The check covers only namespaces you may see.
  • Counts in the templates read 0 only while kube-state-metrics demonstrably reports the namespace. A broken collector does not look like a healthy zero.
  • A refused query says which check refused it, instead of drawing an empty chart.