Cilium WiFi cluster stability tuning¶
Updated 2026-05-07: kensan-lab has since moved to wired LAN as the primary network. WiFi is kept as a fallback, so the relaxed values documented here (
leaseDuration: 60s, etc.) are still in effect (they give breathing room during a fallback failover). This document is kept as historical reference from the WiFi-only era.
On a homelab cluster connecting over WiFi (the RPi 5 fleet on wlan0, the Bosgame M4 Neo on wlp3s0), latency spikes frequently time out lease renewals and API calls. kubernetes/network/cilium/values.yaml relaxes several defaults to accommodate this.
API client rate limit¶
Raise Cilium's client rate limit against the API server. At the default (qps=10 / burst=20), a WiFi spike causes throttling, and the reconcile burst after a Pod recovers then makes the apiserver unreachable.
L2 announcement lease timing¶
The LoadBalancer IP's lease renewal was frequently throwing context deadline exceeded over WiFi, so these were relaxed step by step until landing here:
Trial history¶
| Setting | Result |
|---|---|
| Cilium chart default (15s / 10s / 2s) | Lease renewal failures every few minutes → LoadBalancer IP unreachable |
| 30s / 20s / 4s | Improved, but a timeout still occurred roughly once every few dozen minutes |
| 60s / 40s / 8s (current) | Stable |
Trade-off¶
Extending renewDeadline means a longer delay before IP take-over if the lease-holder Pod crashes. At homelab scale (4 nodes), that's well within tolerance. On a production-grade cluster, staying closer to the default is recommended.
Related¶
- Cilium update strategy
- Known issue: Cilium L2 lease instability over WiFi (also documented in CLAUDE.md's Known Issue section)