An in-place restore onto a different volume than the one backed up is
unsafe: an incremental (CBT) restore computes deltas against a different
volume lineage, and even a full restore would patch and write into an
unrelated volume. Verify the existing PVC is bound and still bound to
the PV recorded at backup time before any side effect, on both the CSI
data mover path (using the backed-up PVC's volume name) and the file
system path (using the PVC-to-PV mapping from the backup volume info).
The PV comparison is skipped for namespace-mapped restores, where the
target PVC is necessarily bound to a different PV (the documented
cross-namespace clone-and-restore workflow).
Signed-off-by: chlins <chlins.zhang@gmail.com>
Check the target PVC is not used by any active pod before any side
effect, on both the CSI data mover path and the file system path.
The in-use semantics align with the pvc-protection controller:
terminal-phase pods don't block, terminating pods block with a wait
hint. On the file system path, pods gated by this restore's
restore-wait init container (identified by the restore UID in its args,
and not yet terminated) are exempted: they must mount the PVC for the
node-agent to restore the data and cannot write to the volume until the
PodVolumeRestores complete. Leftover pods, controller-recreated pods,
and pods gated by a different restore still block.
Signed-off-by: chlins <chlins.zhang@gmail.com>
Add "IncrementalBytes" field to status of DataDownload and PVR to indicate data transferred by the incremental restore
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
* Update CRDs and CLI to support in-place restore (#10038)
Update CRDs(Restore, DataDownload, PodVolumeRestore) and restore create CLI to support in-place restore
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
* Update Kopia(filesystem) uploader to support incremental and deleteExtraFile during restore (#10066)
Update Kopia(filesystem) uploader to support incremental and deleteExtraFile during restore
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
* Update Restore Exposer and PVC CSI to support in-place restore (#10104)
1. Update Restore Exposer to support exposing with existing PV for in-place restore
2. Update PVC CSI RIA to continue the restore process for in-place restore
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
* Update Block uploader to support increase restore (#10244)
Update Block uploader to support increase restore
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
* Update Exposer to recreate the target PV if the volume mode is different with the restore PVC (#10257)
Update Exposer to recreate the target PV if the volume mode is different with t
he restore PVC
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
* Preserve PVC selected-node annotation via carrier annotation for in-place restore
For in-place volume data restore, the existing PVC is deleted and
recreated. For StorageClasses with the WaitForFirstConsumer volume
binding mode, losing the volume.kubernetes.io/selected-node annotation
could let the scheduler place the recreated workload Pod in a different
zone than the original PV, leaving it stuck in ContainerCreating.
Instead of relying on RestoreItemAction execution order (the generic
PVC RIA unconditionally strips the selected-node annotation), the PVC
CSI RIA now captures the annotation from the existing PVC right before
deleting it and carries it on the target PVC via the Velero-internal
restore.velero.io/inplace-restore-selected-node annotation. The restore
engine translates the carrier back to the Kubernetes annotation after
all RestoreItemActions have run and always strips the carrier so it
never lands on the cluster.
This makes the behavior independent of RIA ordering: the Kubernetes
annotation is stripped by default on every path (including when the
target PVC does not exist and Velero falls back to provisioning a new
PVC), and preservation only happens when the CSI RIA explicitly
captured a value from the existing PVC.
Signed-off-by: chlins <chlins.zhang@gmail.com>
* Update the control path to make the in-place incremental restore with block data mover work E2E (#10410)
Update the control path to make the in-place incremental restore with block data mover work E2E
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
---------
Signed-off-by: Wenkai Yin(尹文开) <yinw@vmware.com>
Signed-off-by: chlins <chlins.zhang@gmail.com>
Co-authored-by: chlins <chlins.zhang@gmail.com>
* Report a measured zero incremental instead of erasing it
A CBT incremental with an exactly zero delta -- nothing changed since
the parent -- was reported identically to a backup that moved the whole
device. `velero backup describe --details` printed only
"Moved data Size (bytes): 3221225472" with no incremental line, and
status.incrementalBytes was absent, for a run that transferred nothing.
The best possible CBT outcome displayed as the worst, and was
indistinguishable from a genuine full, a whole-device fallback, or a
backup predating incremental accounting.
The zero was being erased twice. Besides the API status fields,
datapath.BackupResult also carried omitempty, and that struct crosses a
JSON boundary from the data mover pod to the controller (see
micro_service_watcher.go), so the value was destroyed before the
controller could persist it. Every uploader always reports a figure
there, so 0 internally always means "transferred nothing" -- dropping
omitempty is sufficient and correct for that hop.
The API fields move to *int64 rather than just dropping omitempty. The
field shipped in v1.18.0-v1.18.2, so backups exist whose stored volume
info has no incrementalSize at all; with a plain int64 those unmarshal
to 0 and would render "Incremental data Size (bytes): 0", a false claim
of a perfect incremental on a run that never measured one. nil means not
measured, a pointer to 0 means measured zero. Both fields already carry
+optional, so the generated CRD schema is unchanged and no regeneration
is required.
Display gates relax from > 0 to != nil in all three places, including
volumesByPod.Add, whose signature takes *int64 now; the restore describer
passes nil, which is correct since restores measure no incremental.
Verified live: the same zero-delta scenario that reported <none> now
reports 0 and renders "Incremental data Size (bytes): 0", while an older
backup described with the new client still correctly prints no
incremental line at all.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
(cherry picked from commit 6c7aa9d588f6d5eab134d4ce19c92b838f45557c)
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
* gofmt: fix import ordering in backup_test.go
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
* Regenerate CRDs for IncrementalBytes pointer type
make update-crd was missed in the original commit. Regenerated with
the pinned controller-gen v0.16.5 to avoid unrelated version-annotation
churn across other CRDs.
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
* Add changelog for #10309
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
* Address review: make IncrementalBytes a pointer to preserve backward compat
Per Lyndon-Li's review on #10309: dropping omitempty on the plain int64
field breaks compatibility with a data mover from release-1.17 or
earlier that predates IncrementalBytes and never writes the key -- the
new controller would unmarshal a zero value ("nothing transferred")
instead of recognizing the field is simply absent ("not measured").
Switch to *int64 with omitempty restored:
- an old mover's omitted key unmarshals to nil ("not measured")
- a current mover's genuine zero still serializes the key, unmarshaling
to a non-nil pointer to 0 ("measured zero")
- nonzero values work exactly as before
- an old controller can still unmarshal a numeric value from a new mover
pkg/controller/data_upload_controller.go and pod_volume_backup_controller.go
assign the wire-struct field directly to their already-*int64,omitempty
CRD status field instead of re-wrapping it with ptr.To, since both are
now the same pointer type.
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
* Fix CI: update marshal-fail test assertions for IncrementalBytes pointer
Both backup_micro_service_test.go files hardcoded the %v-formatted
zero-value BackupResult struct in an error-message assertion. Now that
IncrementalBytes is *int64, its zero value prints as <nil> instead of 0.
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
---------
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* refactor: use k8s.io/api well-known label constants
Several well-known Kubernetes label strings were hardcoded across the
codebase instead of using the constants already exported by
k8s.io/api/core/v1, which is an existing dependency:
"kubernetes.io/hostname" -> corev1api.LabelHostname
"kubernetes.io/os" -> corev1api.LabelOSStable
"topology.kubernetes.io/zone" -> corev1api.LabelTopologyZone
The local kube.NodeOSLabel and zoneLabel consts, which duplicated the
upstream values verbatim, are now defined in terms of the upstream
constants rather than repeating the literal. Both are kept: NodeOSLabel
is exported and referenced from four packages alongside NodeOSLinux and
NodeOSWindows, which have no upstream equivalent, and zoneLabel sits
beside the deprecated-label fallback it is compared against.
No functional change - every replacement is a constant with an identical
value.
Signed-off-by: Harshit saini <harshitsaini1188@gmail.com>
* Add changelog for #10279
Signed-off-by: Harshit saini <harshitsaini1188@gmail.com>
* Cover the selected-node path in createRestorePod
TestCreateRestorePod only exercised selectedNode == "", so the branch
that pins the restore pod to a node was never executed. Add a case with
a selected node and assert the resulting pod carries the hostname label
in its node selector.
Signed-off-by: Harshit saini <harshitsaini1188@gmail.com>
* Also use constants for the arch and deprecated zone labels
Extends the same replacement to the two remaining well-known labels
raised on the issue:
"kubernetes.io/arch" -> corev1api.LabelArchStable
"failure-domain.beta.kubernetes.io/zone" -> corev1api.LabelFailureDomainBetaZone
zoneLabelDeprecated in item_backupper.go was the last local const still
repeating a literal that upstream already exports, so the zone pair now
reads consistently against k8s.io/api. The deprecation note upstream
applies to the label itself, not the constant; Velero reads that label
deliberately as the fallback for PVs created before the topology labels
existed.
Signed-off-by: Harshit saini <harshitsaini1188@gmail.com>
---------
Signed-off-by: Harshit saini <harshitsaini1188@gmail.com>
* Fix PodVolumeBackup metadata loss on fs-backup timeout
When a backup hits the fs-backup timeout, WaitAllPodVolumesProcessed
returned nil because PVBs were only collected from the indexer in the
done branch of the select. This discarded all PVB metadata including
already-completed PVBs, making their data unrestorable.
Move the PVB collection loop to run after the select so tracked PVBs
are always persisted regardless of timeout.
Fixes#9986
Signed-off-by: Shubham Pampattiwar <spampatt@redhat.com>
* Add changelog for PR #9995
Signed-off-by: Shubham Pampattiwar <spampatt@redhat.com>
* Filter non-completed PVBs in hasPodVolumeBackup
After preserving tracked PVBs on timeout, non-completed PVBs (in-progress
or with no snapshot ID) would cause hasPodVolumeBackup to return true,
leading the restore to skip the original PV and dynamically re-provision
it without any data to restore from.
Only match PVBs that are Completed with a valid SnapshotID.
Signed-off-by: Shubham Pampattiwar <spampatt@redhat.com>
* Add unit tests for hasPodVolumeBackup phase filtering
Verify that hasPodVolumeBackup only matches PVBs that are Completed
with a valid SnapshotID, and rejects in-progress, failed, or
empty-snapshot PVBs.
Signed-off-by: Shubham Pampattiwar <spampatt@redhat.com>
---------
Signed-off-by: Shubham Pampattiwar <spampatt@redhat.com>
Change errors.Cause to errors.Is, because github.com/cockroachdb/errors
New() function create a error with error stack with depth 1, but
github.com/pkg/errors's New() function create error with no depth.
Signed-off-by: Xun Jiang <xun.jiang@broadcom.com>
* Fix UT failures caused by client-go version bump.
* Some modifications to enhance the UT stability.
* Fix UT errors: non-constant format string in call to ...
* Fix linter issues.
Signed-off-by: Xun Jiang <xun.jiang@broadcom.com>
Use typed error approach: Make GetPVForPVC return ErrPVNotFoundForPVC
when PV is not expected to be found (unbound PVC), then use errors.Is
to check for this error type. When a matching policy exists (e.g.,
pvcPhase: [Pending, Lost] with action: skip), apply the action without
error. When no policy matches, return the original error to preserve
default behavior.
Changes:
- Add ErrPVNotFoundForPVC sentinel error to pvc_pv.go
- Update ShouldPerformSnapshot to handle unbound PVCs with policies
- Update ShouldPerformFSBackup to handle unbound PVCs with policies
- Update item_backupper.go to handle Lost PVCs in tracking functions
- Remove checkPVCOnlySkip helper (no longer needed)
- Update tests to reflect new behavior
Signed-off-by: Tiger Kaovilai <tkaovila@redhat.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
chlins
Wenkai Yin(尹文开)andGitHub
R4mboandGitHub
lyndon-liandGitHub
harshit sainiandGitHub
Lyndon-Li
Krishna AwasthiandGitHub
Shubham PampattiwarandGitHub
Xun Jiang
Scott Seago
Matthieu MOREL