Host-originated flows, remote-node, cilium health and DNS :53 were ~74% of the non-reply events on a 12k-event sample and none of them map to a workload card; streaming them all kept hubble-relay at ~400m CPU. They are now blacklisted in the GetFlows request, so agents drop them before the relay. The UI's "Служебный" toggle had nothing left to show and is gone; the legend says what isn't collected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3.6 KiB
cluster-map
Topology map of the Project Chimera Deckhouse cluster.
- Left: inventory tree in two views. Namespaces groups objects by namespace and kind; Ноды shows what is scheduled on each node. Search matches names, images and ingress hosts.
- Centre: the selected namespace(s) or node as compact cards inside namespace boxes. Solid arrows come from configuration (Ingress → Service → workload by label selector); dashed animated arrows are real traffic from Hubble Relay over the chosen window (5 min / 15 min / 1 h), red when Cilium dropped some of it. Peers from other namespaces are drawn as dashed "ghost" cards. Kubelet probes, DNS, cilium health and node-to-node traffic are not collected (see Backend). Traffic refreshes in place every 30 s without relayout; peers that would need new cards wait for "Обновить".
- Right: details for the selected object — overview, links (config and traffic, clickable), live YAML, pod logs.
Workloads under non-native controllers (kruise Advanced DaemonSet behind ingress-nginx, database operators, bare Jobs) are included and named the way Hubble names them, so traffic attaches to the same cards.
Backend
Go, client-go in-cluster config. The graph is built from informer caches
(watch-backed, managedFields stripped), so /api/graph costs milliseconds
instead of ~0.9 s of List calls.
Traffic: each Cilium agent keeps only 4095 flows in its ring buffer — at a
few hundred flows/s that is ~10 s of history, so reading the buffer on
request can't answer "last 15 minutes". A background follow stream from
hubble-relay.d8-cni-cilium.svc:443 (TLS verified with the public part of
Deckhouse's hubble CA at /etc/hubble-ca/ca.crt; relay does not ask for a
client certificate) feeds per-minute buckets for the last hour. Counts are
distinct connections (5-tuples), not Hubble events — one connection
produces several events. new = opened by TCP SYN inside the window,
dropped = tuples with a DROPPED verdict. After a restart the collector
only knows what it has seen since; coveredSec says how much of the
window that is.
Infra traffic is filtered at the agents (infraFilters in flows.go):
host-originated flows (kubelet probes, hostNetwork pods), remote-node,
cilium health and DNS :53. On a 12k-event sample they were ~74% of
non-reply events, none of them attach to a workload card, and streaming
them cost hubble-relay ~400m CPU.
| Endpoint | What |
|---|---|
/api/graph |
nodes + config edges + pod placement |
/api/flows?namespaces=a,b&window=300 |
aggregated traffic touching those namespaces |
/api/resource?kind=&namespace=&name= |
live YAML (managedFields stripped) |
/api/pods?id=wl/<ns>/<kind>/<name>&namespace= |
pods of a workload |
/api/logs?namespace=&pod=&container= |
last 500 log lines |
/healthz |
liveness/readiness |
Frontend
Plain JS, no build step: static/index.html, app.css, app.js.
Cytoscape.js draws, elkjs lays out in two passes (each box on its own with
its unconnected groups packed in rows, then the boxes against each other).
Both libraries are vendored in static/vendor/ so the page works without a
CDN. Theme follows prefers-color-scheme.
Build and deploy
R=git.project-chimera.ru/fanchji/cluster-map/cluster-map-app
docker build -t $R:vX . && docker push $R:vX
docker inspect --format '{{index .RepoDigests 0}}' $R:vX
The Deployment pins the image by digest. Anonymous pull works (public Gitea
user), no imagePullSecret. Manifests: 03-kubernetes/apps/cluster-map/ in
project-chimera, mirrored to k8s-manifests for Argo.