Files

kubernetes/apps/miroir-system/miroir/app/helmrelease.yaml

1 ---

2 # yaml-language-server: $schema=https://k8s-schemas.home-operations.com/helm.toolkit.fluxcd.io/helmrelease_v2.json

3 apiVersion: helm.toolkit.fluxcd.io/v2

4 kind: HelmRelease

5 metadata:

6 name: miroir

7 spec:

8 chartRef:

9 kind: OCIRepository

10 name: miroir

11 interval: 1h

12 upgrade:

13 disableWait: true

14 values:

15 agent:

16 peerFence: true

17 autoEvictAfter: 1h

18 autoDiskfulAfter: 72h

19 drbd:

20 alExtents: 6007

21 groupSnapshots:

22 enabled: true

23 monitoring:

24 dashboards:

25 enabled: true

26 grafanaOperator:

27 enabled: true

28 matchLabels:

29 dashboards: grafana

30 podMonitor:

31 enabled: true

32 prometheusRule:

33 enabled: true

34 replicaCount: 2

72h auto-diskful: cut put_ldev trigger surface while unattended

why ·ok L17

Raised from 10m on 2026-09-01, before a multi-week unattended period. Diskless-to-diskful conversion is a full resync into an attached backing — the same replica-transfer class that LINBIT/drbd#137 names as the put_ldev assertion trigger (there via `toggle-disk --migrate-from`), and both node wedges here (2026-08-17 k8s-0, 2026-09-01 k8s-0) latched during exactly such transfer storms. With ~29 KEDA scale-to-zero apps, a 10m threshold converted legs continually as consumers woke and settled; at 72h a settled consumer just keeps paying the network hop, which costs latency, not a kernel.

Show the rest

This is deliberately a mitigation, not a cure — the medic DaemonSet handles the wedges that still happen. Restore 10m only when the underlying DRBD bug is fixed upstream or someone is around to press power buttons again.

2026-09-0101M1DYCRSY0M4E5PVHRC15NABF

Upgrades must not wait on the agent DaemonSet

why ·ok L12

With helm's default wait, any upgrade that lands while a node is down times out on "DaemonSet/miroir-agent status: InProgress" and rolls back — observed 2026-09-01, when the autoDiskfulAfter change failed to apply because k8s-2 was NotReady. On this cluster that is a standing hazard, not an edge case: nodes go down precisely when miroir is being fixed (DRBD wedges, medic resets), and Renovate auto-merges chart bumps with nobody watching. A rollback loop during a wedge would fight the recovery. The manifests are applied either way; the DaemonSet converges whenever the node returns, and miroir-config's dependsOn still gates on the release itself being made.

2026-09-0101M1DYYJJEERBBZWZR9JTN3VG0