- Shipped
- November 16, 2026 at 5:49 PM UTC
- Author
- Kamo
- Commit
- 8fc0a4f
Timers on k3m1 (02:00, coordinator) and k1m1 (02:05) back up the cluster into **************** the same layout as the first manual backup (11-15-2026): - k8s objects + Secrets, newest scheduled etcd snapshot (read-only copy), rke2 state, Forgejo dump, all database dumps, NATS/Scylla/Qdrant/ RateStack MySQL files, mail, PBX, apps, host config, MinIO (objects, raw drive with versions, IAM). - Space: estimate = 125% of the largest of the last 3 recorded sizes (df growth per run); while df used + estimate > 80%, delete the oldest MM-DD-YYYY folder. The newest complete backup is always kept. - Cluster safety: read-only; sdc reads capped per device in the unit, streams pause on IO pressure, sdc steps start only when no CI build runs; kubectl-exec dumps are slowed, never frozen; unchanged static data is hard-linked from the previous backup; 12 h deadline + one retry. - fb-store on k3m1 is the only writer/deleter on the drive and refuses to write when it isn't mounted. Installed by hand (install.sh), like yugabytedb/: wiring it into deploy-services would edit .forgejo/ and force a full reconcile.
