- Deploy coturn on k8s with MetalLB shared IP (10.0.20.200)
- Normal pod networking (no hostNetwork), runs on any node
- 100 relay ports (49152-49252), port 3478 for STUN/TURN signaling
- Shared secret auth for time-limited TURN credentials
- For F1 streaming WebRTC NAT traversal
- UI and API: 1 → 2 replicas for zero-downtime during restarts/crashes
- Celery worker: Recreate → RollingUpdate strategy
- Celery beat: unchanged (Recreate, singleton scheduler)
- Move f1 from Cloudflare proxied to non-proxied DNS
- Set WEBAUTHN_RPID/ORIGIN for f1.viktorbarzin.me domain
- Add NFS volume at /mnt/main/f1-stream for persistent session/stream data
- Enable headless browser extraction (HEADLESS_EXTRACT_ENABLED=true)
- Reduce replicas to 1 (file-based sessions don't work across replicas)
New skill: music-assistant-librespot-wrong-account
- Documents fix for Spotify playback failing with "librespot does not support
free accounts" when cached credentials point to wrong Spotify account
- Includes step-by-step solution: find container, inspect cache, clear and restart
Updated: home-assistant skill with Music Assistant addon details for ha-sofia
Fixes HA mobile app 403 when embedding Grafana dashboards - the webview
blocks third-party cookies needed by Authentik forward auth. Grafana
already has anonymous Viewer access enabled, so forward auth is not
needed. Also adds grafana_admin_password variable and explicit resource
limits to prevent ResourceQuota issues during rolling updates.
Two skills extracted from multi-user k8s access implementation:
- authentik-oidc-kubernetes: 6 gotchas for Authentik OIDC + kube-apiserver
- kubelet-static-pod-manifest-update: full restart cycle for static pod changes
- Change portal image to viktorbarzin/k8s-portal:latest (Docker Hub)
- Add k8s-portal to cloudflare_non_proxied_names
- Add k8s_users with viktor admin entry to terraform.tfvars
- Add skill_secrets variable to moltbot module with HA tokens and
Uptime Kuma password as container env vars
- Install Python packages (requests, caldav, icalendar, uptime-kuma-api)
in init container with PYTHONPATH for main container access
- Update all skills to use python3 directly instead of ~/.venvs/claude
venv path that doesn't exist in the container
- Remove hardcoded Uptime Kuma password from skill, use env var
- Add RBAC module (modules/kubernetes/rbac/) with admin, power-user,
and namespace-owner roles, API server OIDC flags, and audit logging
- Add self-service portal (modules/kubernetes/k8s-portal/) SvelteKit app
with kubeconfig download and setup instructions
- Configure Alloy to collect audit logs from kube-apiserver
- Add Grafana dashboard for Kubernetes audit log visualization
- Configure Authentik OIDC provider with groups scope mapping
- Wire up k8s_users and ssh_private_key variables through module chain
- Convert setup-project and extend-vm-storage from standalone .md
to directory-based SKILL.md format with YAML frontmatter
- Add symlink in moltbot init container to expose Claude skills
at ~/.openclaw/skills/ for auto-discovery by OpenClaw
- Update CLAUDE.md skill path references
Add MySQL datasource and 15-panel dashboard for DNS analytics:
queries over time, response codes, top domains/clients, response
times, blocked/NxDomain domains. Enable Grafana dashboard sidecar
for auto-provisioning dashboards from ConfigMaps.
Single template regex in the viktorbarzin.lan block catches ALL search
domain expansion junk (*.com.viktorbarzin.lan, *.cluster.local.viktorbarzin.lan,
etc.) instead of needing separate server blocks per pattern. Legitimate
single-label queries (idrac.viktorbarzin.lan) fall through to Technitium.
DNS queries were going through Traefik's IngressRouteUDP, replacing
real client IPs with Traefik pod IPs (10.10.169.150) in Technitium logs.
Changed Technitium DNS service from NodePort to LoadBalancer with
externalTrafficPolicy: Local, removed dns-udp entrypoint and
IngressRouteUDP from Traefik, and updated CoreDNS to forward .lan
queries to Technitium's LoadBalancer IP directly.
- Alloy: bump memory limits from 64Mi/128Mi to 256Mi/768Mi — pods were
OOMKilled at 128Mi, steady-state usage is ~400-450Mi per node
- iDRAC Redfish Exporter: add explicit priority_class_name to resolve
conflict between Kyverno priority injection and default priority: 0
- Set explicit devvm SSH public key for cloud-init (was empty, breaking SSH access)
- Remove hourly cron that restarted all registry containers, which wiped the
in-memory blobdescriptor cache and caused low pull-through cache hit rates
Add kubernetes_config_map for CoreDNS to the technitium module, with a
template block for cluster.local.viktorbarzin.lan that returns NXDOMAIN
immediately. This prevents ndots:5 search domain expansion from flooding
Technitium with ~66k/day junk queries (e.g.
redis.redis.svc.cluster.local.viktorbarzin.lan).
Also enabled saveCache on Technitium so the DNS cache persists across
pod restarts.
Four layers of noisy-neighbor protection using existing tier system:
- PriorityClasses (tier-0-core through tier-4-aux)
- LimitRange defaults auto-generated per namespace tier
- ResourceQuotas auto-generated per namespace tier
- PriorityClassName injection on pods via Kyverno mutate
Custom quota overrides for monitoring and crowdsec namespaces
which exceed the default tier quotas.
Adds check #14 that queries Uptime Kuma API for application-level
monitor status, complementing the kubectl-level checks with HTTP/ping
health data. Reports down monitors by name with PASS/WARN/FAIL thresholds.
