Upgrading

Version-to-version migration steps, newest first. Only releases that need you to do something appear here; any version not listed upgrades by bumping the chart.

miroir is pre-1.0, so breaking changes land in minor versions (0.9.0, 0.10.0, …). Read the section for every version you are crossing, oldest first. Upgrading 0.8.x → 0.10.x means doing 0.9.0 and then 0.10.0.

Check before you upgrade

helm template (or a Flux dry-run) against your new values catches the chart-side failures below without touching the cluster. None of these migrations move data or need volume downtime.

Every upgrade: keep the CRDs in step

The chart ships its CRDs in crds/, and Helm applies that directory only on install, never on upgrade. An upgraded chart running against last release’s CRDs fails in a quiet way: the API server prunes spec fields the old schema does not know, the apply succeeds, and the controller or agent then complains about configuration you can plainly see in your values. So every upgrade starts with the CRDs.

Flux can do it from the chart automatically, but not by default; upgrade.crds must be set (install.crds: Create is already the default):

apiVersion: helm.toolkit.fluxcd.io/v2
kind: HelmRelease
spec:
  install:
    crds: Create
  upgrade:
    crds: CreateReplace # default is Skip; required

Plain Helm has no automatic path; apply the new chart’s CRDs before upgrading:

helm show crds oci://ghcr.io/home-operations/charts/miroir \
  --version <new-version> | kubectl apply --server-side -f -

0.11.x → 0.12.0: the health-monitor sidecar is gone

CSI v1.13 removed the VolumeCondition alpha API the driver used to report replication health, replacing it with dedicated health RPCs (ControllerGetVolumeHealth, ControllerListVolumeHealth, NodeGetVolumeHealth). The driver answers those instead, and no longer advertises the VOLUME_CONDITION capability.

csi-external-health-monitor-controller has not adopted the new RPCs and refuses to start against a driver without VOLUME_CONDITION, so the sidecar and its sidecars.healthMonitor values are gone from the chart. Nothing else consumed the old condition.

Only if you turned it on

sidecars.healthMonitor.enabled defaults to false, so most installs have nothing to do. If yours sets it, remove the whole sidecars.healthMonitor block from your values; leaving it fails the render with a message pointing here.

The PVC events the sidecar raised have no replacement on most clusters: kubelet’s volume-health manager ships in Kubernetes 1.37+ behind the alpha CSIVolumeHealth feature gate (default off), and external-health-monitor has no release that speaks the new RPCs. The miroir_volume_* metrics carry the same signals (split-brain, disk failed, degraded replication) and the chart’s starter alerts already fire on them, so alert there instead. See Monitoring.

0.10.x → 0.11.0: node topology becomes MiroirNode CRs

The storage configuration moves out of the miroir chart entirely: the node topology becomes MiroirNode custom resources (or a MiroirNodeGroup materializing them from a label selector) and, together with the StorageClasses and VolumeSnapshotClasses, is now plain manifests you apply and version yourself; the miroir chart installs only the driver. The CRD schema is what validates the topology, and kubectl apply rejects unknown fields outright. Alongside the move:

• Pool options are grouped under a per-backend block (lvmthin, zfs, or loopfile) instead of prefixed flat keys.
• The per-node setup Job is gone. The agent has always run the same pool provisioning at startup; a misconfigured pool now surfaces in the agent log and the MiroirNode status instead of a failed install hook.
• Canonical spellings are required: volBlockSize uppercase (4K, not 4k), compression lowercase (lz4, not LZ4). Earlier releases folded case; the CRD validates instead.

Existing volumes, their replicas, and their data are untouched. This is a migration of the configuration surface only; do not recreate any PVC, MiroirVolume, or pool.

1. Update the CRDs

As above: upgrade.crds: CreateReplace for Flux, or helm show crds ... | kubectl apply --server-side -f - for plain Helm.

The CRD update must come first

Doing this first is what makes the rest of the upgrade safe: the new schema refuses the partial topology writes 0.10 agents make while the rollout is in flight (see step 3). Skip it and the old schema instead prunes the new pool fields from your manifests. The prune is quiet, and the agents then fail on configuration you can plainly see in your files.

2. Move the storage configuration to manifests

The nodes, storageClasses, and volumeSnapshotClasses values become plain manifests. For nodes, each entry becomes a MiroirNode object of the same name (the old values are the CR spec, reshaped): pools becomes a list of named entries, and each pool’s options move under its backend’s block:

0.10.x values	MiroirNode manifest
nodes.<n> (map key)	metadata.name
nodes.<n>.zone	spec.zone
nodes.<n>.address	spec.address
nodes.<n>.autoEvict	spec.autoEvict
nodes.<n>.pools.<p> (map key)	spec.pools[].name
...pools.<p>.backend	(implied by the block below)
...pools.<p>.device	...pools[].lvmthin.device
...pools.<p>.thinPoolSize	...pools[].lvmthin.poolSize
...pools.<p>.zfsDataset	...pools[].zfs.dataset
...pools.<p>.zfsCompression	...pools[].zfs.compression
...pools.<p>.zfsVolBlockSize	...pools[].zfs.volBlockSize
...pools.<p>.baseDir	...pools[].loopfile.baseDir

There is no backend field any more: the block that is present IS the backend, so exactly one of lvmthin/zfs/loopfile is required even when it has nothing to say: an lvmthin pool whose VG already exists still writes lvmthin: {}. A homogeneous fleet can become one MiroirNodeGroup instead of per-node objects (Quickstart shows the layout).

Each storageClasses entry becomes a standard StorageClass manifest: the chart’s knobs map to parameters (the full table), and what the chart used to default now needs writing: volumeBindingMode: WaitForFirstConsumer, allowVolumeExpansion: true, and the storageclass.kubernetes.io/is-default-class annotation where isDefault was used. volumeSnapshotClasses entries become VolumeSnapshotClass manifests (driver + deletionPolicy).

Before (0.10 miroir values):

nodes:
  k8s-0:
    zone: rack-1
    pools:
      default:
        backend: lvmthin
        device: /dev/disk/by-partlabel/r-miroir
        thinPoolSize: 400g
storageClasses:
  - name: miroir-replicated
    replicas: 2

After (plain manifests):

apiVersion: miroir.home-operations.com/v1alpha1
kind: MiroirNode
metadata:
  name: k8s-0
spec:
  zone: rack-1
  pools:
    - name: default
      lvmthin:
        device: /dev/disk/by-partlabel/r-miroir
        poolSize: 400g
---
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: miroir-replicated
provisioner: miroir.home-operations.com
volumeBindingMode: WaitForFirstConsumer
allowVolumeExpansion: true
parameters:
  miroir.home-operations.com/replicas: "2"
  csi.storage.k8s.io/fstype: ext4

Apply the MiroirNode manifests before upgrading the driver chart, so the rolled agents boot straight into storage mode. Your cluster already has one MiroirNode per storage node (the 0.10 agents created them to publish pool capacity); kubectl apply over them just works: the spec becomes yours, the status stays the agents’. Declaring the fleet as a MiroirNodeGroup instead? The existing MiroirNodes carry no group label, so the group reports them as Conflict and touches nothing (a direct MiroirNode always wins). Hand them over explicitly, and the group then converges each spec to its template:

kubectl label miroirnodes --all \
  miroir.home-operations.com/node-group=<group-name>

The classes need one extra step. The 0.10 miroir release’s manifest still lists your StorageClass/VolumeSnapshotClass objects, and a Helm upgrade deletes objects that dropped out of the manifest by name; the one thing it respects is a helm.sh/resource-policy: keep annotation on the live object. Shield them before step 3:

classes="$(kubectl get storageclasses -o jsonpath='{range .items[?(@.provisioner=="miroir.home-operations.com")]}storageclass.storage.k8s.io/{.metadata.name} {end}')
$(kubectl get volumesnapshotclasses -o jsonpath='{range .items[?(@.driver=="miroir.home-operations.com")]}volumesnapshotclass.snapshot.storage.k8s.io/{.metadata.name} {end}' 2>/dev/null)"
kubectl annotate $classes helm.sh/resource-policy=keep --overwrite

After step 3 they are plain, unowned objects: the manifests you authored above are their source of truth from then on (apply them over the live objects, and drop the keep annotation if you like: kubectl annotate $classes helm.sh/resource-policy-). MiroirNodes need no shielding: the 0.10 chart never rendered them. Clusters tracking unreleased main are the one exception (a development-window chart briefly rendered MiroirNodes); there, extend the annotate command with $(kubectl get miroirnodes -o name).

Loopfile users: also set the driver chart’s agent.loopfileBaseDirs to every loopfile.baseDir in use: the agent pod’s hostPath mounts are pod spec, which the driver chart cannot derive from objects it does not render.

3. Upgrade the driver chart and let the agents roll

Remove nodes, storageClasses, and volumeSnapshotClasses from the miroir chart’s values (it fails fast with a pointer at this page if any is still set) and run the upgrade as usual. While the agent DaemonSet rolls node by node, the not-yet-rolled 0.10 agents keep trying to write their old, partial pool topology into the MiroirNode spec; the new CRD schema rejects those writes. That is by design (it stops an old agent from wiping the pool configuration you just applied), and the visible cost is small: a node’s pool-capacity heartbeat pauses until its agent rolls, so status.observedAt on some MiroirNodes goes stale for the duration of the rollout, and 0.10 agents log rejected MiroirNode updates. Both clear on their own as the rollout completes.

4. Verify

kubectl get miroirnodes
kubectl get miroirnode <node> -o yaml

spec.pools on every storage node should show the per-backend blocks from your configuration (lvmthin.device, zfs.dataset, …). If a block is missing, the CRD update in step 1 was skipped and the old schema pruned it: apply the CRDs and then your manifests again.

Also breaking, but unlikely to affect you

• The --nodes-config flag and --mode=setup are gone, along with the per-node setup Jobs and their ServiceAccount. Only custom agent.extraArgs or tooling that watched the setup Jobs would notice.
• Topology edits no longer roll the pods (the ConfigMap checksum annotation is gone). The controller follows MiroirNode changes live, and each agent restarts itself when its own pool spec changes. This is the new intended behavior, not a regression.
• Two MiroirNodes sharing a replication address no longer keep every component from starting. The conflict is reported as an AddressConflict condition (plus a Warning event) on the offending nodes, which are excluded from new placement until it is resolved.
• helm uninstall no longer destroys volume data by default. The pre-delete hook that deletes every MiroirVolume/MiroirSnapshot is now rendered only when uninstall.confirmation is set to yes-really-destroy-data; see Uninstall. If your teardown automation relied on the old behavior, set the confirmation.
• The controller pod now tolerates node.kubernetes.io/unreachable for 5 seconds instead of Kubernetes’ 300 default, so a dead node stops provisioning for seconds rather than five minutes (unreachableNodeTolerationSeconds restores the old value if you want it).
• The deprecated flat MiroirNode.status.capacityBytes / status.allocatedBytes / status.metaUsedPercent fields are gone from the schema, along with the controller’s fold of them into the default pool. They existed for the 0.9→0.10 rollout skew; 0.10 agents already publish (and read) only status.pools. Anything scraping the flat paths directly has been broken since 0.10; use status.pools[*].

0.9.x → 0.10.0: named storage pools

Each node’s storage config moves under a named pool. Your existing single pool must be adopted as the pool named default: MiroirVolume replicas and StorageClasses written before this release carry no pool reference, and they all resolve to default.

1. Nest each node’s storage under pools.default

zone and address stay node-level. Everything else moves: backend, device, zfsDataset, zfsVolBlockSize, zfsCompression, baseDir, and thinPoolSize.

Before:

nodes:
  k8s-0:
    zone: rack-1
    backend: lvmthin
    device: /dev/disk/by-partlabel/r-miroir

After:

nodes:
  k8s-0:
    zone: rack-1
    pools:
      default:
        backend: lvmthin
        device: /dev/disk/by-partlabel/r-miroir

The chart fails fast on the flat shape, so a missed node is a template error rather than a broken cluster. Under Flux the HelmRelease reports failed and nothing is applied.

2. Upgrade

flux reconcile helmrelease miroir -n miroir-system --with-source

Existing volumes, snapshots, VGs (vg-miroir), datasets, and loopfile directories are reused as they are, with no data migration and no volume downtime. During the rollout an old agent and a new controller can briefly disagree on the MiroirNode status shape; placement treats those stats as unknown (the same as a cold cluster) until the DaemonSet finishes rolling.

3. Optional: add a second pool

nodes:
  k8s-0:
    pools:
      default:
        backend: lvmthin
        device: /dev/disk/by-partlabel/r-miroir
      fast:
        backend: lvmthin
        device: /dev/disk/by-id/nvme-Micron_7450_MTFDKBA800TFS_XXXX
  # k8s-1, k8s-2 identical
storageClasses:
  - name: miroir-replicated
    replicas: 2
  - name: miroir-replicated-fast
    replicas: 3
    pool: fast

Each agent creates the new pool’s VG and thin-pool at startup. New pools get vg-miroir-<pool>; the default pool keeps vg-miroir, which is why step 2 needs no data migration.

Also breaking, but unlikely to affect you

• The agent/setup --lvm-vg and --lvm-thinpool flags are gone; VG naming derives from the pool name. The chart never set them; only custom agent.extraArgs would notice.
• MiroirNode.spec.backend and the flat status.capacityBytes, status.allocatedBytes, status.metaUsedPercent fields are replaced by per-pool lists. Anything scraping those CR fields directly needs the new paths.
• Pool metrics gain a pool label. The chart’s own alerts and dashboard are updated; your own recording rules or dashboards matching an exact label set are not.
• Do not remove a pool from nodes while volumes still reference it. Their reconciles and deletions fail loudly (storage pool "x" is not configured on this node) until the pool returns or the volumes are gone.

0.8.x → 0.9.0: RWX is opt-in

Serving ReadWriteMany is now an explicit operator decision. Gateway pods run privileged in the release namespace, and anyone who can create a PVC could previously cause one to be spawned just by requesting RWX.

Set this before upgrading, not after

If you serve any RWX volumes, set gateway.enabled: true in your Helm values before you upgrade.

gateway:
  enabled: true

Without it, running gateway pods keep serving until their next restart, but the controller stops reconciling them, the gateway RBAC is removed (so a restarted gateway cannot read its volume), and new RWX PVCs are rejected.

If you do not use RWX, no action is needed: the default is false, and the gateway ServiceAccount, RBAC, PodMonitor, and export alert group are simply not installed.

Should you upgrade without the flag and only then notice, setting gateway.enabled: true and reconciling is enough: RWX rejection is FailedPrecondition, so a pending PVC provisions on the external provisioner’s next retry. No PVC recreation is needed.

See ReadWriteMany (RWX) for what enabling it entails.