Backing Up a Windows VPS: VSS Shadow Copies, wbadmin and Offsite Targets

Backup advice for Linux hosts translates badly to Windows. rsync and borg have no equivalent for open database files, VSS writers, or an NTFS journal, and a file-copy backup of a running Exchange or SQL Server instance produces a set that looks complete and restores into a corrupt state. The Windows-native answer is a stack: VSS for consistency, wbadmin (or Windows Server Backup) for orchestration, and an offsite target for the copy that survives the provider’s datacentre.

This article walks the whole chain for a single VPS: how shadow copies produce a consistent point in time, how to script wbadmin so the job is verifiable, and how to get the result onto storage that is not the same hypervisor that runs the machine. It assumes you already have a host; if not, our Windows VPS plans include the snapshot options that this design should sit on top of, not replace.

Snapshots are not backups

A hypervisor snapshot records the virtual disk at a moment in time and is stored on the same storage array as the disk it captures. It is excellent for rollback after a bad patch. It is useless if the account is cancelled, the host is deleted, the storage array fails, or ransomware encrypts the volume and then waits for the snapshot retention to roll over.

The distinction to keep in mind: a snapshot protects against change, a backup protects against loss. Both are worth having, and neither substitutes for the other. The design below produces the second kind.

How VSS makes a live volume copyable

Volume Shadow Copy Service coordinates three parties: the writer (SQL Server, Exchange, the registry), the provider (the Microsoft software shadow copy provider by default), and the requester (whatever is taking the backup). The writer flushes and freezes its state, the provider creates a copy-on-write differential of the volume, the writer is released, and the requester reads from the differential while production continues.

The differential is why the shadow needs somewhere to live. On a small volume with heavy write load, an undersized shadow storage area fills and the copy is aborted mid-backup. Check available shadow storage before trusting a schedule:

vssadmin list shadowstorage
vssadmin list writers
vssadmin list providers

list writers is the diagnostic that matters most. Every writer should report State: Stable and Last error: No error. A writer stuck in Waiting for completion will hang the backup until it times out, and it usually means a service is mid-operation — reboot, or restart the specific service, before retrying.

vssadmin list writers | Select-String -Pattern 'Writer name|State|Last error' | ForEach-Object { $_.Line.Trim() }

Raising the shadow storage limit is a one-liner, but do not go to unlimited on a small system volume — cap it at a percentage you can spare. Ten percent of a 100 GB volume is enough for a backup window of a few minutes on a typical file server workload.

vssadmin resize shadowstorage /for=C: /on=C: /maxsize=10%

wbadmin: what to back up, and what it cannot

wbadmin has two modes worth knowing. Volume backup captures whole volumes and can restore them or mount them for file-level recovery. System state captures the registry, boot files, SYSVOL, Active Directory and the certificate store — everything needed to rebuild a server role, but not your data.

wbadmin start backup -backupTarget:E: -include:C: -allCritical -quiet
wbadmin start systemstatebackup -backupTarget:E: -quiet

wbadmin get versions
wbadmin get status

Note what wbadmin will not do. It cannot write to a network share as a bare UNC path in the default configuration — it wants either a local volume or a mapped drive with credentials passed in, and even then the service account context matters. It does not deduplicate, so a weekly full of a 100 GB volume is 100 GB of target space every week. And it does not verify durability: a successful job exit code means the files were written, not that the target is readable.

A practical schedule on a single VPS is a nightly volume backup to an attached data disk with a seven-day rotation, plus a weekly system state backup retained for a month. Anything longer-lived is handled by the offsite tier below, which copies the wbadmin output rather than re-running it.

Scripting a nightly job that reports failures

A backup that fails silently is worse than no backup, because it removes the instinct to check. Wrap the job so the exit code is examined and the result is written where monitoring can see it.

$log   = "C:\Backup\logs\wb-$(Get-Date -f yyyyMMdd).log"
$start = Get-Date
wbadmin start backup -backupTarget:E: -include:C: -allCritical -quiet *> $log
$code  = $LASTEXITCODE
$stamp = Get-Date

Add-Content C:\Backup\logs\status.csv ("{0},{1},{2},{3}" -f $stamp, $code, ($stamp-$start).TotalMinutes, $env:COMPUTERNAME)
if ($code -ne 0) {
  Write-EventLog -LogName Application -Source 'WbadminNightly' -EventId 4001 `
    -EntryType Error -Message "wbadmin failed with $code, see $log"
}

Register it with Task Scheduler running as SYSTEM with highest privileges, and give the task a hard time limit so a hung VSS writer cannot block the next night’s run. Also enable the “stop if the task runs longer than” option, and check that previous-instance policy is set to Stop Existing rather than running a second copy in parallel.

schtasks /create /tn "Wbadmin Nightly" /sc daily /st 01:30 /ru SYSTEM /rl HIGHEST ^
  /tr "powershell -NoProfile -File C:\Backup\wb-nightly.ps1" /f

The status CSV is the useful artefact. Grafana, Zabbix or a five-line PowerShell check on another host can read the last row and alert on a non-zero code or a stale timestamp — which catches the failure mode where the task never ran at all.

Getting the copy off the hypervisor

Everything above still lives on storage controlled by the same provider account that controls the VPS. The offsite tier is the one that matters in the scenario you are actually planning for. Three options, in ascending order of operational burden:

Option          Mechanism                    Cost          Restore speed
--------------  ---------------------------  ------------  -------------
Object storage  rclone / AWS CLI to S3/B2    Low           Minutes-network
Second VPS      robocopy or rclone pull      Low-mid       Fast
NAS over VPN    SMB to on-prem device        Fixed         Slowest

For most single-server deployments the object storage option is the right default. rclone handles S3-compatible endpoints, does not care that the source is Windows, and gives you a retention policy and integrity checks for free. Push the wbadmin output directory nightly, and keep the tool configuration in a file that is itself part of the backed-up system state.

rclone copy E:\WindowsImageBackup remote:wv-backups/WindowsImageBackup `
  --transfers 4 --checkers 8 --log-file C:\Backup\logs\rclone.log --log-level INFO

rclone check E:\WindowsImageBackup remote:wv-backups/WindowsImageBackup --size-only

Alternatively, a second cheap VPS acting as a pull target is a strong pattern because it inverts trust: the backup host reaches in and copies, so a compromised production server cannot delete its own offsite history. A small second Windows VPS also gives you a machine with Windows already installed for the test restore, which is the next and most important step.

Restore testing: the only proof that counts

An untested backup is a hypothesis. The cheapest test that means something is a file-level recovery from the latest set onto a scratch volume, performed on a schedule rather than once at setup.

wbadmin get versions -backupTarget:E: -machine:WEBSRV01
wbadmin start recovery -version:11/19/2026-01:30 -itemType:File ^
  -items:"C:\inetpub\wwwroot" -recoveryTarget:F:\restore-test ^
  -notRestoreAcl -quiet

Restoring a single folder proves the catalog is readable, the shadow set is complete, and the target credentials work. It does not prove the system can boot. For that, do a full volume restore into a spare VM twice a year: boot from Windows Server media, choose Repair, then System Image Recovery, and point it at the offsite copy.

Two gotchas bite during that test. Bare-metal recovery needs matching driver and platform context; an image taken on one hypervisor may fail to boot elsewhere because the storage controller changed. And version mismatches between the recovery environment and the image build can block the restore outright. Both are found in the test, which is exactly why the test exists.

Retention and the cost of keeping history

Retention math is where backup designs quietly become expensive. A grandfather-father-son scheme on a 100 GB volume at 30 daily, 12 weekly and 12 monthly points is roughly 5.4 TB before compression — not a number anyone wants to pay for on object storage without thinking about it.

Reduce it deliberately. Full volume images weekly, incremental file-level copies daily, and monthly archives kept for a year is a defensible baseline for a small business server. Apply a lifecycle rule at the object storage tier so that monthly archives move to cold storage after 30 days. Then measure how long a restore actually takes — RTO is a decision you make now, not at 3 a.m. during an incident.

Retention layer   Frequency   Kept    Target capacity (100 GB volume)
----------------  ----------  ------  -------------------------------
Daily             7           7       0.7 TB
Weekly            4           4       0.4 TB
Monthly           12          12      1.2 TB
Total                                     ~2.3 TB

The last piece is documentation. A backup plan that lives only in the head of whoever set up the Task Scheduler entry is one resignation away from being lost. Write down the schedule, the target paths, the credentials location, the retention rule and the restore procedure, and store a copy of that document offsite along with the backups. When you need it, you will not be in a position to reconstruct it.

Leave a Comment