- Shipped
- November 17, 2026 at 2:22 AM UTC
- Author
- Kamo
- Commit
- 06469fe
Root cause of the 2026-11-15..17 control-plane flaps: k1m1 ran out of page cache (min free 33G -> 7G in two weeks, major faults 7/s -> 210/s), so cache misses became reads on the spinning RAID that also holds etcd, etcd stalled and rke2 lost its leader lease ~75 times. - kubelet **************** on both nodes (allocatable 123 -> ~106 GiB), so pod requests can't claim the host's memory or page cache; applied on the nodes with an rke2 restart - io-guard: systemd IO*Max= resolves to the raw disk and never bites for LVM IO; kamo-io-guard writes io.max for the dm devices on k1m1 (containerd, docker, registry; etcd/apiserver untouched). Hand-installed, see README - CI on k3m1: no IO caps (no etcd there; builds sat at the cap with the disk 3% busy). CPU/memory caps stay; per-step watcher keeps cpu/memory only - traefik: 1 replica; rolling updates still surge a second pod to switch
