The module lifecycle stage: Generally available version

The module has requirements for installation

v0.9.0

Release date: 2026-11-05

Resources of a node deleted from the cluster no longer hang in Terminating, sds-common-scheduler-extender checks all volumes of a Pod together and no longer counts space twice, the module resources record Events, thin pools no longer depend on the enableThinProvisioning setting, which is removed, and the module images no longer carry the reported golang.org/x/net vulnerabilities.

Highlights

Changes in this release:

  • The controller removes the agent’s finalizer from the LVMVolumeGroup, LVMLogicalVolume and LVMLogicalVolumeSnapshot resources of a node deleted from the cluster once missingNodeGracePeriodMinutes has passed, and new alerts report such nodes and resources that stay in Terminating.
  • sds-common-scheduler-extender places a Pod only on a node whose pools hold all its PVCs together, local and replicated, generic ephemeral volumes included, no longer counts the reservation of a volume against itself, and does not reserve space for scheduling simulations.
  • The enableThinProvisioning setting is removed and the ModuleConfig schema moves to version 2; Deckhouse converts an existing ModuleConfig by itself. spec.volumeCleanup of LVMLogicalVolume now works in the SE and SE+ editions.

New features

This release adds:

  • The controller removes the agent’s finalizer from the LVMVolumeGroup, LVMLogicalVolume and LVMLogicalVolumeSnapshot resources of a node deleted from the cluster, so deleting them no longer hangs in Terminating until the finalizer is removed by hand. It does so once the node has been gone for missingNodeGracePeriodMinutes (10 by default, 0 disables the removal) and the agent’s finalizer is the only one left; an LVMVolumeGroup also waits until no volume or snapshot is left on it. A node that is NotReady or has no running agent does not count as gone. An LVMVolumeGroup that is not being deleted is never deleted automatically, and nothing on the node’s disks is touched. Each removal is recorded as an AgentFinalizerReleased Event, and the time the node was first found gone is kept in the storage.deckhouse.io/node-missing-since annotation of its LVMVolumeGroups.
  • New alerts: D8SdsNodeConfiguratorLVMVolumeGroupNodeMissing fires when the node of an LVMVolumeGroup has been gone from the cluster for more than 6 hours, and D8SdsNodeConfiguratorResourcesOfMissingNodeStuck when resources of a deleted node stay in Terminating past the grace period; it names the reason the finalizer is kept.
  • LVMVolumeGroup, LVMVolumeGroupSet, LVMLogicalVolume and LVMLogicalVolumeSnapshot now record a Kubernetes Event when their state changes: a Warning when a resource fails, a Normal one when it becomes ready again. The reason of the Event is the reason of the condition, for example InvalidConditionStates or ReconcileFailed, so kubectl describe shows when a resource broke and when it recovered. For an LVMVolumeGroup the agent also records the changes of VGConfigurationApplied and VGReady, so a spec the agent refuses shows up as a Warning with the reason ValidationFailed even when the group itself stays Ready. A resource that is only being created, resized or deleted records no Warning, and a pass that changes nothing records no Event. The resources are cluster-scoped, so their Events are in the default namespace, and the controller is granted the right to create them there.
  • LVMLogicalVolume now shows the node and the Volume Group of its logical volume in status.nodeName and status.actualVGNameOnTheNode, as LVMLogicalVolumeSnapshot already does, and d8 k get lvmlogicalvolume shows a Node column. Before, the node had to be looked up through the LVMVolumeGroup. The agent fills the fields in when the volume becomes Created; existing volumes get them after the agent restarts with the update.

Improvements

This release improves:

  • LVMVolumeGroupSet follows its LVMVolumeGroup resources as they change: a resource of the set that is deleted is created again right away, and thinPools, blockDeviceSelector or fileDevices edited by hand are set back to the template. Before, this happened only after the set’s spec changed or the controller restarted. The currentLVMVolumeGroupsCount status field no longer counts resources that were deleted, and the new readyLVMVolumeGroupsCount field (the Ready LVG Count column) shows how many resources of the set are in the Ready phase.
  • A thin pool removed from spec.thinPools of an LVMVolumeGroup is now removed from the node once it holds no thin volumes. Before, the agent ignored the removal and the thin pool stayed. While the thin pool holds volumes, the agent keeps it and reports them in the VGConfigurationApplied condition with the reason ThinPoolHasVolumes; the LVMVolumeGroup stays Ready. Only thin pools that were listed in the spec are removed: the agent now tags them with storage.deckhouse.io/lvmVolumeGroupName.
  • Scheduling simulations of the Deckhouse kube-scheduler (filter and prioritize requests with dryRun: true) no longer reserve space in LVM pools and no longer change the space reservation of the simulated Pod, so the descheduler can check target nodes without taking free space from other Pods. sds-common-scheduler-extender scores nodes in a simulation the same way as in real scheduling, including a Pod with several pending PVCs in one pool and an LVM volume group that serves both a local StorageClass and a ReplicatedStoragePool. Simulation requests are logged at Debug on every extender replica, and /stat shows how many filter and prioritize requests were simulations.
  • The names of Volume Groups and logical volumes are now checked against the LVM rules when a resource is created, instead of failing later in vgcreate or lvcreate on the node and leaving the resource Failed. actualVGNameOnTheNode of an LVMVolumeGroup and of the LVMVolumeGroupSet template, spec.thinPools[].name, actualLVNameOnTheNode of an LVMLogicalVolume and actualSnapshotNameOnTheNode of an LVMLogicalVolumeSnapshot are limited to 127 characters from a-z, A-Z, 0-9, +, _, . and -, must not start with a hyphen or be . or .., and the logical volume names must not start with snapshot or pvmove or contain a suffix LVM reserves for its internal volumes, such as _tdata or _tmeta. An LVMLogicalVolumeSnapshot without actualSnapshotNameOnTheNode is rejected if its own name, which the agent then uses, starts with snapshot or pvmove or is longer than 127 characters. spec.thinPools holds at most 256 thin pools.
  • BlockDevice resources of iSCSI and Fibre Channel LUNs get the status.blockdevice.storage.deckhouse.io/scsitarget label naming the target the LUN belongs to: the iSCSI target name or the WWPN of the Fibre Channel target port, read from sysfs; partitions and multipath maps take it from the disks beneath them. A CSI driver can hide every LUN of the targets it owns with a BlockDeviceFilter by this label, including a LUN that has just appeared on the node and that the driver has not discovered yet. A LUN whose target cannot be read gets no BlockDevice until it can.
  • Thin pool support no longer depends on a module setting: the agent loads the dm_thin_pool kernel module on every node it runs on. A node whose kernel has no dm_thin_pool no longer keeps the agent from starting: the thin-volumes-enabler init container logs dm_thin_pool is not available, thin pools cannot be created on this node and the agent starts. A thin pool requested on such a node is reported in the VGConfigurationApplied condition of the LVMVolumeGroup with the reason ThinPoolCreationFailed or ThinPoolReconcileFailed. Before, the init container failed and the agent pod did not start on the node.

Fixes

This release fixes:

  • The agent could crash or act on a half-updated picture of the node: its cache of block devices, PVs and VGs was written by the scanner and read by the controllers without a lock, and printing the cache after each scan walked the list of logical volumes while a controller added or removed one, which the Go runtime stops with a fatal error. Removing a logical volume that a scan had already dropped from the cache panicked the agent. Every access to the cache now takes its lock, a removed LV that is no longer cached is skipped, and a cached LV handed to a controller is a copy.
  • An LVMVolumeGroupSet did not create the LVMVolumeGroup of a node that joined the cluster, or got the label of its nodeSelector, after the set was created: the controller looked at the nodes only when the set or one of its LVMVolumeGroups changed. It now reacts to nodes that are added, removed or relabelled.
  • An LVMVolumeGroup with several thin pools went NotReady when one thin pool could not be created or extended, and the scheduler stopped placing volumes on its other thin pools too. The LVMVolumeGroup now stays Ready with VGConfigurationApplied=False and the reason ThinPoolCreationFailed or ThinPoolReconcileFailed, the failed thin pool is not used, and the others keep receiving volumes. A thin pool that fails while the Volume Group is created no longer stops the creation of the ones after it.
  • An LVMVolumeGroup whose creation failed after pvcreate could not be created again: the Physical Volume left on the device made its BlockDevice not consumable, and validation refused it from then on. A Physical Volume that belongs to no Volume Group is now consumable with the reason OrphanPhysicalVolume, and a Volume Group takes it as it is.
  • The agent removed the Volume Group of an LVMVolumeGroup being deleted even while another controller still held the resource with its finalizer. The agent now waits until the other finalizers are removed and names them in the VGConfigurationApplied condition.
  • The agent did not create a thin pool from spec.thinPools when another Volume Group on the same node had a thin pool with the same name.
  • BlockDevice resources of a node deleted from the cluster, or of a node that lost the labels the agent is placed by, were never deleted: only the node’s own agent deletes BlockDevices, and none was left there. The controller now marks them Missing with the reason NodeNotFound or AgentNotRunning, Unknown at first and True after 10 minutes, and deletes them after missingBlockDeviceRetentionHours (24 hours by default) unless an LVMVolumeGroup references them.
  • sds-common-scheduler-extender checked the space of a pool for each PVC of a Pod separately, local or replicated. A node passed the filter when the pool held each PVC alone but not all of them, and the Pod could land on a node where one of its volumes was not created. The pool must now hold all the Pod’s PVCs placed in it, including local and replicated PVCs on an LVM volume group that serves both a local StorageClass and a ReplicatedStoragePool. The free space score of a node now uses the space left in the pool after all the Pod’s PVCs, local and replicated, are placed.
  • sds-common-scheduler-extender counted the space reserved for a volume as taken by other volumes when it checked the same volume again. Scheduling a Pod with a local PVC lowered its free space score by the PVC size a second time, because the filter step reserves the PVC right before scoring. When a Pod was scheduled again within a minute of the previous attempt, or the sds-replicated-volume controller repeated a placement request for a replica, a pool whose free space was between the volume size and twice the size was rejected. The volume’s own reservation is now left out of the check; reservations of other volumes still count.
  • sds-common-scheduler-extender scored nodes for Pods with replicated PVCs on a compressed scale: nodes differed by at most 2 points of free space, and in Pods with both local and replicated PVCs the replicated pools barely counted. Replicated PVCs are now scored by free space on the same scale as local PVCs. A pool with less free space than the requested size, or of zero size, counts as no free space, the same as a node without the pool. A PVC whose diskful replica is already on the node no longer lowers the node’s free space score, so with equally free pools a node with replicas of some of the Pod’s volumes is not ranked below a node without replicas. A node with replicas of all volumes still gets the top score, and its lead over other nodes now depends on their free space. A node with replicas of only some volumes may lose its lead: with 3 volumes and pools more than about half free, a node with replicas of 2 of them is scored the same as a node without replicas.
  • An LVMLogicalVolume whose deletion waits for a finalizer of another module (for example, storage.deckhouse.io/sds-local-volume-csi) kept Ready=True and the Created phase, and the agent logged the wait only at Debug, so the deletion looked stuck for no reason. The agent still removes the logical volume only after the other finalizers are removed, but while it waits the resource now reports Ready=False with the reason Deleting and lists the finalizers in the condition message, and the agent logs the wait at Info. The phase stays Created, since the logical volume is still on the node.
  • sds-common-scheduler-extender ignored generic ephemeral volumes (spec.volumes[].ephemeral) of a Pod: such a Pod was scheduled without checking whether the node has an LVM volume group of the StorageClass with enough free space, and the volume could fail to be created on the chosen node. The extender now takes the PVC that Kubernetes creates for the volume, <pod name>-<volume name>, and checks it like any other PVC of the Pod.
  • A second LVMVolumeGroup could be created for a Volume Group that another LVMVolumeGroup already described. Both resources became Ready over the same Volume Group, the scheduler extender counted its space twice, the agent moved the storage.deckhouse.io/lvmVolumeGroupName tag of the Volume Group between them on every pass, and deleting the resource the tag named at that moment removed the Volume Group the other one still described. The new lvg-validation webhook rejects the creation of such a resource and names the existing one. If one gets into the cluster anyway, the agent serves only the resource the Volume Group is tagged for, or the older one, and reports the other with VGConfigurationApplied=False and the reason ValidationFailed.
  • A label of a BlockDevice that the agent derives from its status (kubernetes.io/hostname, kubernetes.io/metadata.name, status.blockdevice.storage.deckhouse.io/*) could be changed by hand, and the change held until the next device event on the node, because the agent rewrites them only on a discovery pass. A wrong value sent the device to the wrong blockDeviceSelector of an LVMVolumeGroup or LVMVolumeGroupSet. The new d8-sds-node-configurator-blockdevice-label-guard ValidatingAdmissionPolicy rejects changing, removing or adding such labels for everyone except the agent and names the keys; other labels stay free to set.
  • sds-common-scheduler-extender rejected a node for a Pod that only mounts a local volume already on it when the thin pool was overcommitted on paper, for example after a volume was restored outside the scheduler: the bound PVC was checked against the pool’s free space, which was negative, and the Pod stayed Pending. A bound PVC that does not grow no longer claims space in the pool.
  • spec.thinPools[].size of LVMVolumeGroup and LVMVolumeGroupSet and spec.size of LVMLogicalVolume are declared as integer or string, and the API server accepted a number such as size: 1073741824, but the module’s components read the field as a string only. One such object made every list of its resource fail to decode, so the agent, the controller and the scheduler extender stopped seeing all LVMVolumeGroups or LVMLogicalVolumes of the cluster. A number is now read as its decimal text, the same as the quoted string.
  • In the default block device discovery, through lsblk, the agent took mount points from its own container instead of the node. A disk mounted on the node whose filesystem signature was missing from the udev database, for example after wipefs, was reported as consumable and could be taken into an LVMVolumeGroup. The agent now takes mount points from the node’s mount table (/proc/1/mountinfo) in both discovery modes, and such a BlockDevice is not consumable with the reason Mounted.
  • When a Volume Group was not imported because its storage.deckhouse.io/lvmVolumeGroupName tag names the LVMVolumeGroup of the same Volume Group on another node, the message in the agent’s log suggested only the remedy for shared storage: move the LVMVolumeGroup to this node. That advice is wrong for a disk cloned from another node, so the message now also covers a clone: change the UUID of the Volume Group with vgchange -u and remove the tag with vgchange --deltag.
  • On a node without thin-provisioning-tools, such as the Debian and Ubuntu cloud images, LVM activated a thin pool without checking its metadata and printed WARNING: Check is skipped, please install recommended missing binary /usr/sbin/thin_check!, so a pool with damaged metadata went into service. The module now installs thin_check, thin_dump and thin_repair under /var/lib/deckhouse/sds/bin, with the loader and libraries they need under /var/lib/deckhouse/sds/lib/pdata_tools, and the agent’s LVM uses them on every node whatever the node has installed. A thin pool whose metadata fails the check is no longer activated: LVM reports Check of pool ... failed (status:...). Manual repair required!.
  • An LVMVolumeGroupSet kept the LVMVolumeGroup of a node deleted from the cluster in its status and counted it among its current and Ready resources. Such a resource now drops out of the status; it is not deleted.
  • spec.volumeCleanup of LVMLogicalVolume now works in the SE and SE+ editions. The agent of these editions refused every cleanup method with volume cleanup is not supported in your edition and left the LVMLogicalVolume being deleted in the Failed phase with its logical volume on the node. The agent now erases the volume before removing it in every edition except CE, and an LVMLogicalVolume left in this state is erased and deleted once the agent of the new version starts.
  • A BlockDevice whose update took it out of a BlockDeviceFilter (for example, a grown LUN that a filter by size hides) was deleted only on the next discovery pass, which waits for a udev event, and stayed available to LVMVolumeGroup resources until then. Now it is deleted in the same pass.
  • An LVMLogicalVolume with a percent size (for example, size: 50%) was not extended after its Volume Group or thin pool grew if the free space held the growth but not the whole new size of the volume: the agent compared the full target size with the free space, in which the current size of the volume is already counted as used. A 100% or 100%FREE volume could never grow. The volume stayed Failed with unable to extend LV, err: not enough space. The agent now compares only the growth: with the free space of the Volume Group for a thick volume and, for a thin volume, with the space left in the thin pool under its allocationLimit.
  • An LVMLogicalVolumeSnapshot created without spec.actualSnapshotNameOnTheNode stayed Pending forever: the API server rejected the agent’s update that adds its finalizer, because the field had to be non-empty. The field may now be empty or absent, and the snapshot on the node is then named after the resource. Once set to a non-empty value, the field stays immutable.
  • A disk with a partition table and no partitions in it, for example a leftover GPT header, was reported as a consumable BlockDevice, and pvcreate refused it only when an LVMVolumeGroup was built on it. Such a device is now not consumable with the reason HasPartitionTable, in both discovery modes. The agent does not wipe it; the FAQ explains how to check and clear the table.
  • An LVMLogicalVolume whose logical volume vanished from the node right after it was created, for example when another LVM writer overwrote the Volume Group metadata, could not be deleted: the agent kept calling lvremove for a volume that did not exist, failed with Failed to find logical volume and released the finalizer only after it was restarted. Deleting a volume created moments before the agent’s next scan could also crash the agent. The agent now checks with LVM that the logical volume is gone and releases the finalizer.

Security updates

Security updates in this release:

  • The Go dependencies of the agent, the controller, the scheduler extender, the webhooks and the module hooks are updated: golang.org/x/net from v0.60.0 to v0.62.0, closing CVE-2026-78661 and CVE-2026-97045. The module images are built on base images v4.0.7.

Breaking changes

Changes that affect backward compatibility:

  • The enableThinProvisioning module setting is removed, and the ModuleConfig schema moves to version 2. The agent loads the dm_thin_pool kernel module on every node, so thin pools can no longer be turned off with enableThinProvisioning: false.

Upgrade notes

Before upgrading, note the following:

  • The enableThinProvisioning setting is removed and the ModuleConfig schema moves to version 2. Deckhouse converts an existing ModuleConfig and drops the setting whatever its value. With enableThinProvisioning: false the agent now loads dm_thin_pool as well. The release notes of v0.5.16 said the setting had been removed; it had not, and it kept working until this release.
  • Before the upgrade, remove enableThinProvisioning from the sds-node-configurator ModuleConfig if it is set there. Deckhouse converts a ModuleConfig only when its status.version is filled in. If the ModuleConfig was created before the module was loaded and its status.version is empty, Deckhouse validates the settings against the new schema without converting them, refuses enableThinProvisioning as a forbidden property, and the ModuleRelease stops in the Pending or Suspended phase with a ModuleConfig validation error.
  • The command of the thin-volumes-enabler init container changes in all clusters, so the agent DaemonSet restarts its pods once.

Docs

Documentation changes:

  • The FAQ explains how to make storage on a new node come up by itself: put a label into /var/lib/node_labels on the node before it joins the cluster, and select it in dataNodes.nodeSelector of the module and in an LVMVolumeGroupSet.
  • The FAQ explains how allocationLimit of a thin pool works: how status.thinPools[].availableSpace is calculated, what is refused once the limit is reached, and what happens when the limit is changed. The descriptions of allocationLimit and availableSpace in LVMVolumeGroup and LVMVolumeGroupSet now say the same.
  • The FAQ section on changing the UUID of Volume Groups when cloning virtual machines now explains that the clone of a node whose Volume Group was created by the module also carries the storage.deckhouse.io/lvmVolumeGroupName tag, so the agent on the clone does not import the Volume Group, and that the tag has to be removed with vgchange --deltag after vgchange -u.
  • The FAQ explains what happens to the LVMVolumeGroup, LVMLogicalVolume, LVMLogicalVolumeSnapshot and BlockDevice resources of a node deleted from the cluster and how to clean them up.
  • The FAQ explains how CSI drivers hide their LUNs from the module with BlockDeviceFilter resources: the rules by WWN, by model and by target, how the scsitarget label value is built, and the limitations.
  • The requirements for a device to get a BlockDevice resource no longer contradict the description of creating a partition on a block device: a device with partitions or other devices built on it gets no BlockDevice of its own (unless it is an LVM Physical Volume), each of its partitions gets one if it meets the requirements, a partition with a PARTUUID needs no serial number, and a partition table without partitions leaves the disk not consumable with the reason HasPartitionTable.
  • The requirements for a device to get a BlockDevice resource now explain why the disk the OS is installed on and its partitions get no BlockDevice resources with a usual layout, including a root on LVM, and the FAQ explains how to give the module the free space on that disk: a new partition, the root Volume Group handed over with the storage.deckhouse.io/enabled=true tag, or file-backed devices.
  • The FAQ describes how to disable the module: which modules to disable first, in which order to delete LVMVolumeGroupSet and LVMVolumeGroup resources to free the disks, and what stays on the nodes and in the cluster.
  • The description of status.usedSize of LVMLogicalVolumeSnapshot and of its UsedSize column now says what the value is: the space the snapshot maps when it is created, blocks shared with the origin volume included. It does not grow as the origin is written to and is not the space that deleting the snapshot would free.

v0.8.2

Release date: 2026-10-26

The agent no longer offers a path of a multipath map as a BlockDevice, steps around multipath maps whose every path has failed and reports them with metrics and alerts, and its DaemonSet is ready for the Pod Security Standards enforcement of DKP 1.79.

Highlights

Changes in this release:

  • A path of a multipath map never becomes a BlockDevice, even when udev has not marked it mpath_member, so an LVMVolumeGroup can no longer be built over a raw path of a LUN that already carries a volume’s data.
  • A multipath map whose every path has failed is now treated like a map with no paths: the agent’s LVM commands do not read it, it is not offered as a BlockDevice, and the new alerts D8SdsNodeConfiguratorUnreadableMultipathMaps and D8SdsNodeConfiguratorLVMScanAbandoned tell the operator such maps exist and how to remove them.
  • The module ships a SecurityPolicyException for the agent DaemonSet, so its Pods keep being created once DKP 1.79 starts denying Pods that violate the Pod Security Standards in system namespaces.

New features

This release adds:

  • The SecurityPolicyException sds-node-configurator in the d8-sds-node-configurator namespace describes what the agent needs beyond the restricted profile: privileged containers running as root, the host network and PID namespaces, the spc_t SELinux type, its hostPath volumes and the metrics hostPort. The agent Pods are labeled security.deckhouse.io/security-policy-exception: sds-node-configurator. Both are rendered only when the admission-policy-engine CRDs are present in the cluster. All three containers of the agent now drop all capabilities, as the restricted profile requires; the privileged ones keep their access, because the container runtime ignores the drop list for them.
  • The module ships its first Prometheus rules. D8SdsNodeConfiguratorUnreadableMultipathMaps fires when a node has held a multipath map with no usable path for 10 minutes; D8SdsNodeConfiguratorLVMScanAbandoned fires when a pvs, vgs or lvs run that the agent left behind at its deadline has been running for 5 minutes. The alert texts explain the consequences, such as a CSI volume that cannot be detached from the node, and the steps to remove the maps.

Improvements

This release improves:

  • New agent metrics: sds_node_configurator_unreadable_multipath_maps{node} counts the multipath maps the agent keeps LVM away from, sds_node_configurator_lvm_scans_abandoned_total{node,command} counts the LVM scans left running at their deadline, and sds_node_configurator_lvm_scans_outstanding{node,command} is 1 while such a scan is still on the node. The agent log names the unreadable maps, as multipath -ll shows them, at Info level when it first sees them and whenever the set changes.

Fixes

This release fixes:

  • A path of a multipath map (a device listed in /sys/block/dm-N/slaves of a map whose dm/uuid starts with mpath-) could be published as a consumable BlockDevice. udev marks a path mpath_member only if multipath claimed it when it appeared, and with find_multipaths yes it does not claim a path that appeared before the LUN’s WWID was registered, which is the order in which a CSI driver connects a volume. The agent now reads the paths of every multipath map from sysfs and never makes a BlockDevice of them.
  • A multipath map that still listed its paths but had every one of them failed, with queue_if_no_path on, blocked any reader just like a map with no paths, yet the agent’s filter caught only the latter: lvs hung on such a map for hours holding neighbouring maps open, multipath -f failed with map in use, and a ReadWriteOnce volume could not move to another node. The agent now reads the multipath target status through the bundled dmsetup, which does no I/O to the maps, and rejects a map with no active path in the LVM global_filter and drops it from BlockDevice candidates. The set of such maps is rebuilt right before every scan pass, so a map that lost its last path a moment earlier is not opened once more.
  • An LVM scan (pvs, vgs, lvs) that outlived its deadline was killed, and a scan blocked on an unreadable device turned into a process in the D state that only a reboot removes, still holding the device open. Such a scan is now left running and exits on its own once the device is dealt with; while it lives, the agent starts no other scan of the same kind.
  • The sds-utils-installer init container of the agent logged a failed copy of the node utilities and exited successfully, so the agent started with the utilities of the previous release. It now fails, and the Pod does not start until the copy succeeds.

Security updates

Security updates in this release:

  • The Go dependencies of the agent, the controller, the scheduler extender, the webhooks and the module hooks are updated, closing CVE-2026-78669, CVE-2026-78659, CVE-2026-78660, CVE-2026-78663 and CVE-2026-97032 in golang.org/x/net, CVE-2026-56851 in golang.org/x/text and CVE-2026-63209 in github.com/klauspost/compress. github.com/docker/distribution, which has no fixed release for CVE-2026-85747, is no longer linked into the module hooks.

Upgrade notes

Before upgrading, note the following:

  • Nothing has to be done by hand. The agent Pod template changes on every cluster, so the agent DaemonSet restarts its Pods once.
  • A BlockDevice already created on a path of a multipath map is handled like one whose device left the node: a consumable one is deleted, and one that is not consumable gets the Missing condition and is deleted after missingBlockDeviceRetentionHours unless an LVMVolumeGroup uses it.

Docs

Documentation changes:

  • FAQ: why multipath maps with no usable path and the paths of a multipath map never become BlockDevice resources, which metrics report them, and how to remove such maps with dmsetup message <map> 0 fail_if_no_path and multipath -f <map>.
  • A section on adding capacity on a node used by a LocalStorageClass: the class accepts one LVMVolumeGroup per node, so capacity is added by extending the existing Volume Group through its spec.blockDeviceSelector instead of creating a second LVMVolumeGroup.

v0.8.1

Release date: 2026-10-01

The agent deletes BlockDevice resources that have stayed Missing past a configurable retention, the scheduler extender no longer promises the same free space twice in HA mode, and inactive thin volumes are activated again.

Highlights

Changes in this release:

  • New module setting missingBlockDeviceRetentionHours (24 by default): a BlockDevice whose device has been gone for that long and that no LVMVolumeGroup uses is deleted, so nodes with dynamically mapped LUNs no longer collect Missing resources forever.
  • With two replicas of sds-common-scheduler-extender (HA mode), space reservations are served by one elected replica, so two Pods can no longer be promised the same free space of a thin pool.
  • A thin LVMLogicalVolume deactivated while its thin pool stayed active is activated again by the agent.

New features

This release adds:

  • Module setting missingBlockDeviceRetentionHours (hours, 0 to 8760, 24 by default). A BlockDevice that is not consumable is not deleted when its device disappears: it gets the Missing condition, True after 10 minutes of absence. The agent now deletes such a resource of its node once Missing has been True for the retention, provided no LVMVolumeGroup lists it in status.nodes and its status.lvmVolumeGroupName does not name an existing LVMVolumeGroup, so a Physical Volume lost from under a live Volume Group stays visible. If the device returns, the resource is created again. 0 disables the deletion. The new metric sds_node_configurator_block_devices_missing{node,reason} counts the Missing=True resources that remain; each deletion is logged by the agent.
  • On DKP 1.78 and later, the module’s access roles follow the capability/scope RBAC scheme: the ClusterRoles d8:system-capability:sds-node-configurator:view and d8:system-capability:sds-node-configurator:edit grant read-only and full access to BlockDevice, LVMVolumeGroup, LVMVolumeGroupSet, LVMLogicalVolume and LVMLogicalVolumeSnapshot. Below 1.78 the module keeps the d8:manage:permission:module:sds-node-configurator:view and :edit roles.

Fixes

This release fixes:

  • In HA mode, the two sds-common-scheduler-extender replicas each kept their own space reservations in memory, so scheduling requests answered by different replicas could be promised the same free space and overcommit a thin pool. The replicas now elect a leader over the Lease d8-sds-node-configurator/sds-common-scheduler-extender; only the leader serves filtering, scoring and reservation requests, and the other replica forwards them to it, answering 503 while no leader is reachable.
  • A thin LVMLogicalVolume deactivated on its own (for example, with lvchange -an) while its thin pool stayed active was never activated again, so the volume stayed unavailable. The agent now runs vgchange -ay for its Volume Group as it already did for thick volumes and thin pools. Logical volumes carrying the activation skip flag, which the module never sets, are still left inactive.

Upgrade notes

Before upgrading, note the following:

  • Deletion of Missing resources is on by default. Right after the update, the agent deletes every BlockDevice that has already been Missing=True for more than 24 hours and is not used by an LVMVolumeGroup. To keep them, set missingBlockDeviceRetentionHours to 0 or to a larger value before updating.
  • On DKP 1.78 and later, the d8:manage:permission:module:sds-node-configurator:view and :edit ClusterRoles are replaced by d8:system-capability:sds-node-configurator:view and :edit. Bindings that reference the old role names directly have to be switched to the new ones.

Docs

Documentation changes:

  • The Russian reference for BlockDevice now covers status.conditions and its fields.

Dependencies

Dependency updates:

  • LVM2 (lvm, dmsetup on the nodes): main snapshot → 2.03.42 (changelog)
    • The static lvm and dmsetup the module installs on the nodes now come from the LVM2 2.03.42 release. The previous build pinned a revision but never checked it out, so nodes got an unreleased snapshot of the LVM2 main branch.

v0.8.0

Release date: 2026-09-24

The module’s files on the node move from /opt/deckhouse/sds to /var/lib/deckhouse/sds, the agent stops reading LVM from virtual machine disks, and thin snapshots no longer depend on the dm_snapshot kernel module.

Highlights

Changes in this release:

  • The module’s node tree moves to /var/lib/deckhouse/sds, which stays writable on nodes whose /opt is read-only, and the fileDevicesDirectory default becomes /var/lib/deckhouse/sds/file-devices. Existing spec.fileDevices entries under the old directory keep working where they are.
  • On a node that runs virtual machines, the Volume Groups a nested cluster builds on its disks no longer turn into LVMVolumeGroup and BlockDevice resources of the host cluster.
  • LVMLogicalVolumeSnapshot, restores from a VolumeSnapshot and clones of thin volumes work on nodes without the dm_snapshot kernel module.

Improvements

This release improves:

  • New metrics: sds_node_configurator_published_block_devices and sds_node_configurator_published_devices_directory_present show how many devices the agent keeps out of LVM on each node and whether it found kubelet’s directory of published block volumes there, and sds_node_configurator_legacy_file_devices_directory_lvgs counts the LVMVolumeGroup resources whose spec.fileDevices still name /opt/deckhouse/sds/file-devices; the module log lists their names.

Fixes

This release fixes:

  • On a node that runs virtual machines (DVP), the agent read LVM inside VM disks: a guest cluster’s Volume Groups appeared as LVMVolumeGroup resources and its disks as BlockDevice resources of the host cluster. The agent’s lvm commands now reject every device kubelet has published to a Pod as a block volume, with its partitions and whatever is stacked on it, for any CSI driver. Resources created this way before the update leave Ready and can be deleted; the Volume Group inside the VM is not touched.
  • A Volume Group tagged storage.deckhouse.io/lvmVolumeGroupName=<name> for an LVMVolumeGroup that does not exist in this cluster was imported as a new resource, although only the module in another cluster (for example, inside a virtual machine) writes that tag. It is no longer imported: the agent logs the refusal once and counts it in sds_node_configurator_lvm_volume_group_import_refused_total. A Volume Group handed over with storage.deckhouse.io/enabled=true and no name tag is imported as before, and so is a file-backed one.
  • Creating an LVMLogicalVolumeSnapshot, restoring from a VolumeSnapshot and cloning a thin LVMLogicalVolume failed on nodes without the dm_snapshot kernel module loaded, with snapshot: Required device-mapper target(s) not detected in your kernel. Thin snapshots are now created without that module. The agent still loads dm_snapshot on every node with thin provisioning, and a node where it is missing no longer keeps the agent from starting.
  • The bundled lvm could not load a missing device-mapper kernel module by itself and reported the target as not detected instead. It now calls the node’s /sbin/modprobe.

Breaking changes

Changes that affect backward compatibility:

  • The module no longer installs lvm and dmsetup under /opt/deckhouse/sds/bin, and removes the old copies from nodes. Scripts outside the module that run them from there have to use /var/lib/deckhouse/sds/bin/.

Upgrade notes

Before upgrading, note the following:

  • Nothing has to be done by hand. Once the agent Pod has restarted on a node, the next bashible run removes the files the module used to install under /opt/deckhouse/sds; files of other modules and backing files there are left in place. Scripts that ran /opt/deckhouse/sds/bin/lvm or dmsetup have to switch to /var/lib/deckhouse/sds/bin/.
  • spec.fileDevices[].directory is immutable, so entries created with the old default /opt/deckhouse/sds/file-devices keep naming it, and the agent keeps accepting it. Do not move their backing files by hand; the FAQ describes how to move an entry to the new directory online. Create new entries under the new default. A fileDevicesDirectory set explicitly in the module config is kept as is.

Docs

Documentation changes:

  • FAQ: a section on backing files created before the base directory moved, with a procedure that moves an entry to the new directory using pvmove.

v0.7.5

Release date: 2026-09-14

The scheduler extender now runs on the control-plane nodes, next to the kube-scheduler that consults it.

Highlights

Changes in this release:

  • The sds-common-scheduler-extender Pod is placed on control-plane nodes instead of system ones, and tolerates their taints.

Fixes

This release fixes:

  • The sds-common-scheduler-extender Deployment asked for system nodes and carried the system tolerations, which tolerate neither the control-plane taint nor custom ones, so the Pod could not be placed on a control-plane node. It now selects master nodes — falling back to system nodes in clusters that expose no master role — and tolerates any node, which puts it next to the kube-scheduler that calls it on every scheduling attempt of the modules that register it.

v0.7.4

Release date: 2026-09-11

Controller and scheduler extender metrics are now published through a kube-rbac-proxy, and percent-sized LVMLogicalVolumes are no longer rejected or left in Failed once they reach their requested size.

Highlights

Changes in this release:

  • The /metrics endpoints of the controller and the scheduler extender are no longer reachable without authentication: both bind on 127.0.0.1 and a kube-rbac-proxy publishes them, so a scrape needs get on deployments/prometheus-metrics.
  • A percent-sized LVMLogicalVolume — 100%FREE of a whole volume group, the shape sds-elastic uses for one OSD per disk — is no longer refused at creation and no longer parked in Failed with unable to extend LV, err: not enough space.
  • An LVM scan that reaches a multipath map with no remaining paths no longer blocks the agent, so the node keeps reporting its block devices and its RWO volumes can move to another node.

Improvements

This release improves:

  • Metrics of the scheduler extender are collected for the first time: the manager had always bound the port, but nothing declared or scraped it.
  • utils_commands_execution_count_total and utils_commands_errors_count_total are exported at last. Both were incremented in about forty places in the agent and the controller but never registered, so nothing they counted was visible.
  • All binaries of the module log through one shared logger. The name of the component that emitted a record is now a field of the record (logger) rather than a bracketed prefix inside the message text.

Fixes

This release fixes:

  • LVMLogicalVolume with a percentage in spec.size is computed against a capacity LVM reports as the raw device sum, which is not extent-aligned. Rounding the request up landed one extent past that capacity, so a request for 100% of a volume group was judged not to fit into it: creation was refused, and the extender put an already full-size volume into Failed. Percent sizes are now floored to the extent boundary.
  • An LVMLogicalVolume that had been marked Failed because the group’s free space still lagged behind the just-created LV stayed Failed for the rest of its life, since later passes returned early on a volume that was already at its requested size. Such a volume now moves to Created.
  • The agent could report block device size is invalid for a healthy device: the ioctl output variable lived on the stack and a stack growth during the call moved it, so the size was written to the old address and stayed zero.
  • An LVM scan that reached a multipath map whose paths were all gone blocked in the kernel instead of failing, and every stuck scan held the map open. That stalled the agent, kept multipath -f from removing the map, and left the volume in Node.status.volumesInUse, so it could not be attached elsewhere. Such maps — a mpath- device-mapper UUID with no slaves — are now rejected in the global_filter of every LVM command.

Security updates

Security updates in this release:

  • Anonymous access to the controller and scheduler extender metrics is closed. Previously the endpoints authenticated no one, were reachable from every pod in the cluster, and the ServiceMonitor sent the Prometheus ServiceAccount token to them over plain HTTP.
  • golang.org/x/crypto used by the module hooks is updated to 0.56.0, which closes CVE-2026-56854 (critical, authentication bypass in x/crypto/ssh), CVE-2026-56855 and CVE-2026-78662.

Upgrade notes

Before upgrading, note the following:

  • Nothing has to be done by hand for metrics: the ServiceMonitor shipped with the module already scrapes the proxy over HTTPS with a token. Only a collector of your own that used to read the controller metrics directly over HTTP has to be pointed at the proxy port and granted get on deployments/prometheus-metrics.
  • Log records changed shape, so a filter that matched the component name inside the message text has to match the logger field instead.

v0.7.3

  • Added the Missing condition to BlockDevice: a missing but still occupied device (e.g., part of an LVM Physical Volume) no longer remains forever with a stale status.path. It is first set to Unknown; True is set only if the device is continuously absent for 10 minutes, so a flapping iSCSI path that recovers on its own is not marked. The condition is cleared as soon as the device is found again; the object itself is not deleted to avoid breaking references from LVMVolumeGroup and LVMLogicalVolume
  • Fix: the agent no longer crashes on a cached logical volume without LVM attributes
  • Fix: VPA ceilings for the agent, scheduler-extender, and controller are raised to match measured consumption
  • Fix: the release image is no longer built from a cache with an outdated changelog.yaml — git-mapping dependencies are tied to the install stage where the file is copied
  • Updated base images to v2.1.3 and lib-helm to 1.72.14
  • Documentation: described virtual disk serial numbers
  • Internal build changes for the module

v0.7.2

  • Bugfix: the controller is granted permission to write events (create and patch) - previously the event write was denied by RBAC
  • CVE fixes
  • Internal build changes: CI updated, crds, docs and openapi are shipped in the release image

v0.7.1

  • Correction: one “orphaned” device-mapper device without a disk no longer stops BlockDevice detection on the entire node - previously, because of it, the agent did not create a single BlockDevice on the node, including healthy disks
  • Correction: LVMVolumeGroup with a single PV whose name could not be determined immediately no longer hangs in Pending forever - empty status.nodes post and no retry in the absence of udev events have been fixed
  • Correction: the agent substitutes the serial number inherited from the parent device if direct reading of the serial number fails (multipath participants, crypt/dm devices) - previously such devices were skipped when detecting BlockDevice
  • Optional resources (VerticalPodAutoscaler, PodMonitor/ServiceMonitor) are now enabled based on the presence of a CRD in the API, and not based on the enabledModules list
  • Documentation: fixed duplication of the description key in the Russian CRD LVMVolumeGroup, which broke docs-builder, added the missing status.extentSize description and restored the blockDevice description

v0.7.0

  • Correction: the agent no longer accepts LVM structures inside block devices given to virtual machines as its own - the /dev/loop exception was returned in the LVM global_filter. Fixed the resulting LVMVolumeGroup failures with CacheEmpty and ScanFailed causes - the VG name is no longer considered an identifier, and the LVM warning about a duplicate of someone else’s name is not attributed to our group
  • Added support for LVM on files: spec.fileDevices field of LVMVolumeGroup and LVMVolumeGroupSet. The agent creates a file, connects it as a loop device and uses it as a PV, which gives LVM storage on a node without a free disk; the blockDeviceSelector field has become optional
  • The scheduler (scheduler-extender) takes into account the PVCs listed in the extra-pvcs pod annotation
  • LVMLogicalVolume, LVMLogicalVolumeSnapshot and LVMVolumeGroupSet publish status.conditions; BlockDevice now has a Consumable condition, explaining why the device cannot be used; processing of LVMVolumeGroup conditions has been brought to a general form
  • Correction: a volume created from a snapshot or clone is expanded to the requested size - previously the PVC remained in Pending waiting for a size that never appeared
  • Correction: thin pool set by percentage of VG size no longer gets stuck at VGConfigurationApplied=False due to LVM rounding by one extent
  • Correction: an idle call to lvextend is recognized as harmless on LVM 2.03.16 - otherwise LVMVolumeGroup remained in NotReady
  • Correction: when taking control of an existing VG, blockDeviceSelector no longer remains empty
  • Correction: after pvcreate and vgcreate the agent calls udevadm to update the udev database
  • Correction: when storage-foundation is enabled, the agent loads the dm_snapshot kernel module
  • Removed outdated CRD LvmVolumeGroupBackup, left over from a one-time migration: the new hook removes finalizers from leftovers and removes the CRD itself
  • Added an experimental mode for detecting block devices via netlink, enabled by the ENABLE_NETLINK_BLOCK_DEVICE_DISCOVERY flag
  • Updating base images to v1.3.25, Go 1.26.5 and lib-helm to 1.72.13
  • Fixed vulnerabilities in golang.org/x/net, golang.org/x/sys and golang.org/x/crypto
  • Internal changes in module assembly and CI, the set of e2e tests has been significantly expanded

v0.6.10

  • The general accept rule was removed from LVM global_filter, which led to scanning of non-standard paths (/dev/disk/by-diskseq/*) and duplication of VG names

v0.6.9

  • The agent now ignores “foreign” block devices (rbd, drbd, nbd, loop) when scanning LVM
  • Fixed handling of duplicate VG names: LVMVolumeGroup status is no longer corrupted
  • The blocking of the agent pod launch has been eliminated; added timeouts for LVM commands
  • Improved stability of legacy block device names: matching by serial number or WWN prevents re-creation when name changes
  • Fixed BlockDevice selector validation: the In operator is now the only valid one

v0.6.8

  • Fixed race in scheduler-extender when PVC.Status.Phase is empty: PVC binding decision is now made based on Spec.VolumeName, nodes without a suitable LVMVolumeGroup are always rejected
  • Agent is now also scheduled on sds-elastic data nodes
  • Updating base images and lib-helm to 1.72.0

v0.6.7

  • Update base images, Go 1.25.10 and lib-helm 1.71.12
  • Internal changes in module structure and assembly

v0.6.6

  • Internal changes in module structure and assembly

v0.6.5

  • Update base images, Go 1.25.10 and lib-helm 1.71.11
  • Internal changes to the module assembly

v0.6.4

  • Corrections to the module structure

v0.6.3

  • scheduler-extender: PVCs with a missing StorageClass no longer block pod scheduling

v0.6.2

  • CI changes: DistroPackagesProxy and env proxy in werf, improvements to CVE scans (role_name, checkout)
  • Added user-authz cluster roles in templates

v0.6.1

  • Added detection of mount points via the mountinfo parser in the agent
  • Added locality of replicas to scheduler-extender for replicated volumes
  • Fixed reconciliation races and loops in LVG/LLV controllers
  • ExtentSize was taken into account when comparing in discoverer and fallback 4Mi was added in the LLV controller
  • Removed false marking of LLV as Failed when actualSize > spec.size
  • Fixed double subtraction of Failed LLV from actualSize when calculating free space in scheduler-extender
  • Updating base images and lib-helm

v0.6.0

  • CVE fixes
  • Transition to a common scheduler for SDS
  • Replacing a fixed resize delta with automatic determination of the LVM partition size delta
  • Automatic activation of LVM partitions during volume operations

v0.5.16

  • Added a hook for clearing resources when the module is turned off
  • Removed the enableThinProvisioning setting, we assume that thin provisioning is always enabled
  • Module-sdk update for CVE fix, golang update

v0.5.15

  • Added metrics for LVM volumes
  • Module-sdk update for CVE fix
  • Refactoring manifests

v0.5.14

  • Updated base images and lib-helm versions
  • Set default value for enableThinProvisioning to true
  • Docs improvements
  • Fixed templates security context
  • Added support for HA mode

v0.5.13

  • Updated base images versions

v0.5.12

  • Updated Go version to 1.24.9
  • Updated lib-helm to deckhouse_lib_helm-1.64.1
  • Documentation updates

v0.5.11

  • Fixed volume cleanup for thin volumes in some OSes

v0.5.10

  • Removed obsolete python requirements

v0.5.9

  • Fixed bug with unactivation of logical volumes

v0.5.8

  • Added release notes

v0.5.7

  • Fix sys-fs-cgroup error for containerdv2 support

v0.5.6

  • Added additional mounts for containerd v2 support

v0.5.5

  • CVE fixes

v0.5.4

  • CVE fixes
  • Internal changes for containerd v2 support

v0.5.3

  • Fixed bug preventing LVG status updates when missed/notready nodes are present in the cluster

v0.5.2

  • Removed support for sds-drbd module
  • Added hiding of csi-scsi-generic devices

v0.5.1

  • Technical release, module refactoring

v0.5.0

  • Added ability to wipe data in enterprise versions
  • Multiple documentation fixes

v0.4.6

  • Technical release. Removed “Preview” status from documentation

v0.4.5

  • Numerous documentation fixes
  • Added LVMVolumeGroupSet and LVMLogicalVolumeSnapshots
  • Fixed work with labels and finalizers in some internal objects
  • Added dm_snapshot loading when needed

v0.4.3

  • Fixed crash of BlockDevices migration script when finalizers are missing on resources

v0.4.2

  • Fix for supporting BlockDevice serial numbers up to 63 characters long

v0.4.1

  • Fixed potential issue with missing labels in LvmVolumeGroup->LVMVolumeGroup migration hook
  • Added controller that tracks labels on BlockDevice resources for quick updates of LVMVolumeGroup resources

v0.4.0

  • LvmVolumeGroups resources will be migrated to LVMVolumeGroups
  • Exact BlockDevices lists in LVMVolumeGroups will be migrated to selectors
  • Multiple fixes in controllers and documentation

v0.3.2

  • Fix in sds-health-watcher-controller which could incorrectly handle the status of some nodes

v0.3.1

  • Updated golang to current 1.22.6 to close known vulnerabilities

v0.3.0

  • Added thin provisioning configuration in mc for automatic dm_thin_pool module loading
  • Images moved to distroless
  • Added labels to BlockDevices entity (for future use in BlockDevicesSelectors)
  • Multiple bug fixes and documentation improvements

v0.2.5

  • Added metrics and their collection in Prometheus, health check ports moved to correct ones

v0.2.4

  • Added auto-expansion for thin pools LVM volume groups and LVM logical volumes

v0.2.3

  • Added internal cache for performance improvement
  • Added health and readiness checks to controller
  • Improved resource display in cli
  • Now LVM volume groups are deleted when corresponding k8s resource is deleted, not when annotation is added
  • Added LVM volume group size specification in percentages
  • Added AllocationLimit support in LVM volume group
  • Added support for contiguous volumes in sds-local-volume

v0.2.1

  • Added multipath devices support

v0.2.0

  • Added LVMLogicalVolumeWatcher CRDs
  • Added more test cases
  • Added AreSizesEqualWithinDelta function
  • Fixed resize for thinPools
  • Added tests to LvmLogicalVolumeWatcher controller
  • Fixed naming in LVMLogicalVolume
  • Added node affinity
  • Enhanced LVMLogicalVolume handling with logging, size display, and event processing
  • Statically linked nsenter, lsblk and lvm utils
  • Fixed serial discovery by switching to dynamic lsblk
  • Rewrote bin-copier script in Golang
  • Moved lsblk and its libraries inside the agent image
  • Added age field
  • Implemented stderr filtering for LVM commands
  • Added bench-tests for LVMLogicalVolume controller
  • Changed size type from string to quantity
  • Added parallel reconciliation to lvm_logical_volume_watcher and VG size validation to lvm_volume_group_watcher