Repository navigation
Five pages asked for in #331: a map of the cluster, who talks to whom, what
the workloads are configured to risk, what the cluster costs, and a history
that outlives the window. None of them installs anything into a cluster, and
each says which of its sources it could not read.
Added
-
Topology. A map of what the cluster runs and how it hangs together:
routes and Ingresses into Services, Services into workloads, the
ConfigMaps, Secrets and volumes a workload uses, and the autoscalers and
disruption budgets that govern it. Group it by namespace or by app, find
an object with/, lay NetworkPolicies over it, export it as an image. A
route or a Gateway is toned by the verdict its own page gives, and a kind
the cluster refused is named, with the lines it takes away. -
Traffic. Who calls whom, with request rate, share of errors and
latency where the source measures them. It reads Istio, Hubble, Beyla and
Caretta series from the Prometheus the cluster is connected to, and live
flows from Hubble Relay, dropped ones with their reason and the
NetworkPolicies that select the pod they were bound for. A workload's
page and peek show who calls it and whom it calls. A workload a source
cannot see is "not observed", never "no traffic". -
Audit. Over forty rules across security, reliability, efficiency and
upgrade: what the Pod Security Standards refuse, bindings that hand out
cluster-admin, missing probes and budgets, requests far above measured
use, and API versions a Helm release or a writer still uses that the next
Kubernetes removes. A finding names the field and its value, and a rule
that could not be checked says so instead of passing. Every workload page
has an Audit tab. -
Cost. Per hour, per month and idle, by namespace, workload and node,
from OpenCost's series in the connected Prometheus, or from prices you
enter where there is none. A workload's and a namespace's page carry
their own cost line, and efficiency compares requests with a day of use.
A part without a price is "not priced" and makes the total "at least".
Changed
-
Changes keeps its history. What changed in the cluster is recorded
while the app is connected and kept on this machine for up to seven days
(and up to 20 000 changes a cluster), so it is still there after a
restart; a stretch the app was not watching is drawn as one, and rows let
go early are said to be. Events stand beside changes on one clock, and every object
has a Changes tab. Values of Secrets, ConfigMaps and environment
variables are never kept, only that they changed. What was recorded
while the app could not tell which cluster a context names (a reader who
cannot readkube-system) is kept and says so, since another cluster may
have had the name. Settings › Diagnostics shows what is kept and clears
it. -
Wide tables open faster. Column widths are measured in one pass
rather than once per value.
Fixed
These were in 5.0 too.
-
A cluster made again under the same name is taken without a restart.
When a connected context's entry in the kubeconfig changed (a new
address, certificate or user), the app went on dialling the old one. -
A Prometheus that was still being probed was offered as missing, and one
that did not answer was never asked again; the page now asks again. -
A Service's or a workload's Connections tab did not say that a route's
Gateway names a class no controller has accepted, which the Routes page
did. -
A route attached through a ListenerSet whose Gateway is missing was told
"Gateway <the set's name> does not exist"; it now names the set as found
and its Gateway as the one missing. -
Changes dropped rows past its limit without saying so.
-
A condition's reason and message were drawn as the cluster wrote them; a
private key a controller pasted into one is now withheld there as it is
everywhere else. -
A cluster, namespace, node or key named
constructoror__proto__
is read as any other. A cluster named__proto__lost its alias,
colour and critical mark, so marking it critical did not gate it; one
namedconstructorcould not be switched to; a ConfigMap key of that
name showed as withheld. -
A drain called a disruption budget with no status a blocker; it is now
"not known". -
A route whose parent is a Service was matched to a Gateway of the same
name. -
A kubectl line the app hands out to copy (a list row's
kubectl get,
My Access'sauth can-i, the check beside a refused action) names the
context and the kubeconfig files the app read it from, so it asks the
cluster on screen and not the one a terminal happens to point at.