Перейти к содержимому
Платформа

Платформа

Networking

Cloudflare Tunnel ingress, Tailscale private access, Cilium, and the default-deny posture that ties the cluster together.

На этой странице
Эта страница ещё не переведена, поэтому показана английская версия.

Public traffic reaches the cluster only through an outbound Cloudflare Tunnel. Operator access goes through Tailscale. Inside the cluster, every namespace starts from a default-deny network policy and opens only the paths it needs. No connection gets trust from its network position alone.

The hosting and tailnet decision is recorded in ADR 0004; this page is the operational reference.

The two planes

flowchart LR
    subgraph Public["Public plane"]
        Browser["Visitor browser"]
        Tebex["Tebex webhooks"]
    end

    subgraph Edge["Cloudflare edge"]
        Tunnel[Cloudflare Tunnel]
    end

    subgraph Cluster["Cluster"]
        Cloudflared["cloudflared<br/><i>DaemonSet</i>"]
        Traefik["Traefik<br/><i>networking namespace</i>"]
        Services["Services"]
    end

    subgraph Tailnet["Tailnet (private overlay)"]
        OperatorOn["Operator device"]
        Proxies["Tailscale proxies<br/><i>admin console,<br/>NATS monitoring</i>"]
    end

    Browser --> Tunnel
    Tebex --> Tunnel
    Tunnel --> Cloudflared
    Cloudflared --> Traefik
    Traefik --> Services

    OperatorOn -. WireGuard .-> Proxies
    Proxies --> Services
  • Public plane (Cloudflare): anything that must be reached by a party we do not control, such as broadcasters opening the dashboard and Tebex posting purchase webhooks. Twitch events do not arrive here: ingress opens outbound EventSub WebSocket connections.
  • Private plane (Tailscale): operator access to the admin console, NATS monitoring, and the Kubernetes API.

Cloudflare Tunnel: the public ingress

cloudflared runs as a DaemonSet on every node except the control plane (deploy/operators/cloudflared.yaml). Every ingress rule forwards to the Traefik Service in the networking namespace over TLS. The connector trusts only the Cloudflare Origin CA, and it forces the SNI to the zone apex so routing never depends on which public hostnames exist.

What is exposed

Hostname Purpose
dashboard.itsbagelbot.com The broadcaster console (SvelteKit). The app’s /status, /healthz and /readyz are masked at Traefik.
stats.itsbagelbot.com The public stats page.
commands.itsbagelbot.com, leaderboard.itsbagelbot.com Public commands pages and leaderboards, served by the dashboard app through the wildcard rule.
health.itsbagelbot.com Per-service /status endpoints read by the hosted status page (deploy/k8s/status-routes.yaml).
webhooks.itsbagelbot.com Tebex webhooks to the transactions service.
itsbagelbot.com and *.itsbagelbot.com Everything else on the zone. Traefik owns the routing decision, and unrouted hosts fall back to the marketing site.

status.itsbagelbot.com is the hosted status page itself. Its DNS points at the status provider, so the tunnel does not claim it.

The tunnel configuration ends with a required http_status:404 rule.

Why a tunnel

A tunnel keeps the nodes off the public internet and gives edge TLS termination and DDoS absorption for the public webhook receiver. Tebex gets a stable, branded webhook URL instead of a node address.

Tailscale: the private overlay

The Tailscale Kubernetes operator publishes selected Services to the tailnet through Ingress resources with ingressClassName: tailscale:

Service Where defined
Admin console deploy/k8s/console-admin.yaml
NATS monitoring deploy/messaging/nats.yaml
Traefik dashboard and the Falco UI Allowed by deploy/networking/network-policies.yaml and deploy/observability/network-policies.yaml

The admin console therefore has no public hostname. Operators reach it only from a device on the tailnet.

The Cilium configuration points the Kubernetes API server address at a tailnet address (deploy/cilium/values.yaml). The tailnet policy itself is not stored in this repository.

Inside the cluster

  • CNI: Cilium in native routing mode with kube-proxy replacement, BBR bandwidth management, and BIG TCP (deploy/cilium/values.yaml).
  • Network policies: each namespace carries a default-deny policy plus explicit allows (deploy/*/network-policies.yaml, deploy/k8s/network-policies.yaml). Cross-namespace paths name the source namespace and pod labels.
  • Service-to-service transport: NATS carries all inter-service traffic (ADR 0003). Valkey and NATS use TLS with certificates issued through cert-manager (deploy/cache/certificates.yaml, deploy/messaging/certificates.yaml). The MySQL connection pins the HeatWave CA through DB_CA_CERT. Traefik re-encrypts to the dashboard console, so that hop is TLS end to end.
  • Secrets: the Doppler operator syncs secrets into the namespaces that need them (DopplerSecret resources next to each workload).

What is deliberately not here

  • A LAN ingress. No NodePort or LoadBalancer Services exist in deploy/.
  • A VPN concentrator on the cluster. Tailscale is the VPN; we do not run a second one.

Where to next