Whole-machine backup (bare-metal recovery)
A Back up a whole Windows machine job images every disk Windows needs to boot (or every fixed disk) as a bootable VHDX, kept as a weekly full copy plus nightly changes. Today you restore it as a Hyper-V or Proxmox VM or download the VHDX, and a boot check can start a copy every week to prove it works; restoring onto new hardware from a USB stick comes later.
What it does today, and what it does not #
- It captures the whole machine: one bootable
.vhdxper physical disk, partitions at their original places, with the boot partitions, the partition table and amachine.jsondescribing the firmware, layout and drivers for each restore point. - It restores a restore point as a new Hyper-V VM on a host you choose, as a new Proxmox VM on a cluster node you choose, or as one bootable VHDX per disk downloaded to a Windows machine you choose.
- Physical to virtual works today. A failed server can be running as a VM on any Hyper-V host or Proxmox cluster while its replacement is ordered.
- It proves itself. A boot check restores the newest restore point to a throwaway VM on a schedule, watches it reach the Windows sign-in screen, keeps a screenshot and deletes the VM.
- Not yet: booting a new or wiped machine from a recovery USB stick and writing the image back to its disks (bare-metal restore to new hardware). It is being built; see Recovery USB (coming). Until then, restore to a VM, or attach the downloaded VHDX to one.
- Windows only. The job runs on the Windows agent of the machine being imaged, never on the Cloud Agent.
Recovery USB (coming) #
Restoring onto a physical machine will work like this: you boot the machine from a Caryvane recovery USB, type a one-time code, and the USB writes the restore point onto the machine's disk and makes it bootable. The USB tool itself is not released yet. What exists so far:
- Recovery codes. In the machine's Restore points, Create recovery code… makes a code such as
7KQ2-M9XD-4TPAfor one restore point. It is shown once, lasts 4 hours, 24 hours or 3 days, and needs the same access as restoring the machine to a VM. Creating one is recorded in the audit trail. - What a code allows. It opens that one restore point, read-only. The first machine to use it keeps it; anyone who types it later is refused, as if the code were wrong. The USB downloads the image straight from your storage through links that stop working after a few hours, and never receives your storage keys. Every use, and every refused try with a real code, is recorded.
- Storage. Codes work for images on Amazon S3, Wasabi, Backblaze B2, Cloudflare R2, other S3-compatible storage and Azure Blob. An Azure storage connected without its account key cannot hand out these links; connect it with the account name and key.
- Disks. The replacement disk must be at least as big as the space the original's partitions used; restoring onto a smaller disk by shrinking Windows' partition is not in the first version.
Create the job #
- Jobs → New Job → Back up a whole Windows machine, then choose the machine (only Windows agents are offered).
- Choose which disks (see Which disks below). Disks Windows needs to boot is the right choice for most machines.
- Choose storage that encrypts at rest (see BitLocker). The wizard shows a rough storage estimate from the machine's used space and the version history you keep.
- Schedule. Nightly is the default. The first night copies all the used data on the chosen disks; after that, each night uploads what changed.
Which disks #
| Choice | What it images | Use it for |
|---|---|---|
| Disks Windows needs to boot (recommended) | Every disk holding a volume Windows reports as critical for booting: the system and boot partitions, and the Windows volume. | Most machines. Enough to start the machine again. |
| Every fixed disk | All internal disks, data disks included. USB and removable disks are left out. | A server whose data disks belong in the same restore point. |
| Choose disks | The disks you tick. The agent lists the machine's disks with their model, size, type and drive letters, and greys out any that cannot be imaged, with the reason. | Particular disks on their own schedule, or leaving out a disk that cannot be imaged. |
A job fails if any disk it selects cannot be imaged, so Every fixed disk fails on a server with a dynamic data disk. The disk list warns you. A dynamic data disk needs only a Files backup: Windows cannot boot from it, so an image of the boot disk plus a Files backup of the data restores the whole server.
A job remembers disks by number, and numbers can change when disks are added or removed, so check the job after changing the disks in a machine. If a chosen disk is not found, the run stops and says so.
How each night works #
Each disk is kept as a chain, laid out like a Hyper-V chain: a full base copy once a week, then a small difference layer each night holding only the blocks that changed. A restore point is the base plus the layers up to that night.
- How changes are found. The agent reads the disk through a Volume Shadow Copy snapshot and keeps one hash per 1 MiB block. Each night it hashes the blocks again and uploads only those whose hash changed.
- It reads all the used data every night. Uploads are small, but the reading is not: allow about 40 minutes per TB of used space, depending on the disks. Schedule it for when the machine is quiet.
- What is skipped. Free space,
pagefile.sysandhiberfil.sysare not read or stored; Windows recreates them. - ReFS volumes. Where ReFS does not offer a free-space map, the whole volume is read, free space included. That is slower and the first copy is larger; the run says so.
- A fresh base every week keeps chains short, so a restore never has to apply a long run of layers, and lets older chains be removed as a whole once your version history no longer needs them.
The change journal. On NTFS the agent can use Windows' change journal to read only the files that changed rather than every used block. Windows' own journal on C: is 32 MB, which a busy machine fills in a couple of hours, so by the next night it has wrapped and the whole volume is read again. From agent 1.0.332 a whole-machine job raises the journal of each volume it images to 1 GB (a bigger one is never shrunk) and says so once in the run log, with the fsutil usn createjournal command that undoes it. A volume that would be left with less than 10% or 4 GB free is left alone, and the run log says why. To opt out, set Aurawolf:UsnJournalMaxMB to 0 in the agent's settings (%ProgramData%\Aurawolf\agent.settings.json); any other number is the size in MB.
As a rough guide to storage, each weekly chain is about the used space plus six nights of changes (about 2% of the used space a night on a typical server), and the job keeps as many chains as your version history covers, plus the one being built.
BitLocker: the image holds decrypted data #
A BitLocker volume is read through the snapshot while it is unlocked, so the image contains its data decrypted. Anyone who can read the image can read the disk. Keep the image as well protected as the machine itself: storage access limited to the people who need it, and the storage key kept safe.
Because of this, the destination must encrypt at rest, and Caryvane refuses to save a whole-machine job pointed at storage that does not: Amazon S3, Wasabi, Backblaze B2, Azure Blob, DigitalOcean Spaces, Cloudflare R2, Linode or Vultr object storage. It is the same rule as for Shopify and Xero backups. SFTP, FTP, SMB shares and local folders are not accepted for this job. Other S3 storage (OVH, Scaleway, MinIO and so on) is accepted when its bucket has default encryption switched on: Caryvane asks the bucket when you save, and if it does not, the message says how to switch it on.
A VM restored from the image boots without BitLocker on the restored volumes. Turn BitLocker on again inside it if it is going to stay in service.
When a run refuses, and what to do #
A run never backs up nothing quietly. If a disk cannot be imaged faithfully, the run stops with the reason named in Run History:
| Reason | What to do |
|---|---|
| Dynamic disk | Dynamic disks are not supported. Convert to a basic disk where you can, or back the data up with a Files job and choose the other disks for the image. |
| Storage Spaces | Pooled virtual disks cannot be imaged as physical disks. Back their contents up with a Files job; image the boot disks only. |
| Software RAID, or a volume spanning disks | Not supported. Hardware RAID that Windows sees as one disk is fine. Otherwise use a Files job for the spanned volume. |
| Locked BitLocker volume | Unlock the volume (or set it to unlock automatically) so the snapshot can read it, then run the job again. |
| Disk not found | The disk number has changed or the disk is offline. Check Disk Management and correct the number in the job, or choose Disks Windows needs to boot. |
Restore a machine #
Restore points are listed per machine: Agents → ⋮ on the machine → Restore points…. Each shows the date, the firmware (UEFI or BIOS), the Windows version, the total size, the disks (with the ones Windows needs to boot marked) and how many layers deep the chain is. Each restore runs as a job on the machine doing the work and shows in the Jobs list and the job's run history.
Restore to a Hyper-V VM #
- Restore to a Hyper-V VM… on the restore point.
- Choose the Hyper-V host (an online Windows machine with Hyper-V) and a scratch folder on it, for example
D:\Restores. Optionally name the VM and set its memory (4096 MB by default) and processors (2 by default). - The host downloads the chain into the scratch folder and creates the VM. The VM's generation follows the original machine's firmware: Generation 2 for UEFI/GPT, Generation 1 for BIOS/MBR.
- The VM is created with no network adapter connected and is not started, so it cannot clash with the original machine on the network. Check it, connect a virtual switch when you are ready, and start it.
Download bootable VHDX #
- Download bootable VHDX… on the restore point, then choose the Windows machine and the scratch folder.
- The machine downloads each disk's chain and joins it into one
.vhdxper disk in the scratch folder, ready to attach to a VM, mount in Windows to copy files out, or keep.
Restore to a Proxmox VM #
For a customer with a Proxmox cluster that has the Caryvane agent on its nodes.
- Restore to a Proxmox VM… on the restore point.
- Choose the cluster, the node the VM is created on, and the Proxmox storage for its disks, for example
local-lvmor a Ceph pool (not a Caryvane backup storage). Optionally name the VM and set its memory and processors. - The node's agent creates the VM, allocates each disk on the storage at its original size, and writes the restore point straight into it, reading the backup in place from your storage: nothing is copied to the node first, so the node needs no spare space of its own. Empty space is skipped on thin storage (LVM-thin, ZFS, Ceph RBD with KRBD, directory, NFS, CIFS, Btrfs); on thick LVM every block is written. The VM's firmware follows the original machine's: UEFI (OVMF) for UEFI/GPT, with Secure Boot on only if the original had it, and BIOS (SeaBIOS) for BIOS/MBR.
- The disks are attached as SATA. Windows has its own driver for SATA, so a machine imaged from physical hardware starts without any extra drivers; VirtIO is faster but Windows has no driver for it until you install one. Choose IDE for a very old machine that ran its disks in IDE mode, or VirtIO SCSI only if the machine already has the VirtIO drivers.
- The VM is created with its network card unplugged and is not started, tagged
caryvane-restore. Check it in Proxmox, plug the network in when you are ready, and start it.
The Proxmox storage must keep a quarter free after the disks are written: on thin storage that counts the disks' data, on thick storage their full size. Only a backup storage that cannot be read in place (none a whole-machine backup can use today) has its layers downloaded first, to /var/tmp/caryvane-restore on the node's own disks; never point that at a shared or NFS folder, which running VMs use.
The 25% rule. Every restore refuses to start if it would leave the scratch folder's drive, or the Proxmox storage, with less than 25% free, so a restore never fills the host's disk. Choose a roomier drive or free some space and run it again.
Send a physical machine to a VM with the Caryvane stick (beta) #
A machine with no Caryvane agent, or one that will not start Windows, can be sent straight to a Proxmox or Hyper-V host as a new VM. Boot the machine from the Caryvane stick; its disk is only read. The copy goes over your local network to the Caryvane agent on the host and never passes through Caryvane.
- On the stick, choose Send this machine to a Proxmox host or Send this machine to a Hyper-V host, and the disk if the machine has more than one. A BitLocker disk asks for its recovery key and is sent decrypted.
- The stick shows a code and a QR code. Scan it with your phone (or open the console and choose Approve a Caryvane stick…), and check the machine it describes. The stick needs internet for this moment only.
- Choose the customer, then the host: a Proxmox node with the Caryvane agent, or a Windows machine with Hyper-V and the Caryvane agent. It must be the customer's own or your company's, online, and on a recent agent; hosts that cannot be used are listed with the reason.
- For Proxmox, choose the storage for the VM's disk (for example
local-lvm). For Hyper-V, choose a folder on the host, or leave it empty for Hyper-V's own default. Optionally name the VM and set its memory and processors. - Approve. The host's agent connects to the stick on the local network (the stick listens on TCP port 47192; the host needs no inbound firewall rule) and both prove the approval before anything is sent. The stick accepts that one connection only.
- The host writes the disk as it arrives and makes the VM: on Proxmox, written straight onto the storage, with its disk on SATA and BIOS or UEFI as the machine had; on Hyper-V, as a
.vhdxinfolder\VM name, Generation 2 for a UEFI machine and Generation 1 for a BIOS one. The VM is not started and its network is unplugged. The job shows in the host's run history.
- Fast Startup and hibernation. If Windows was shut down with Fast Startup or hibernated, the copy's hibernation image is cleared on the way, so the VM starts Windows fresh instead of trying to resume on different hardware. The original machine is not changed.
- A machine that had the Caryvane agent. The VM carries that agent's identity. Before connecting the VM's network, delete
C:\ProgramData\Aurawolf\token.jsoninside it (or uninstall the agent); the stick, the VM's notes and the run all say so. - BIOS machines stay BIOS. Turning a BIOS (MBR) machine into a UEFI Generation 2 VM is not part of this.
- One disk per send. Send a second disk the same way, then attach it to the VM.
- The same 25% rule applies to the Proxmox storage or the Hyper-V folder's drive. Approving needs the same access as restoring a machine, and at least one enrolled machine on your account; it is recorded in the audit trail with who approved it, for which customer, and the host.
Boot checks: proving a restore point starts #
A boot check restores the machine's newest restore point to a throwaway VM on a schedule, starts it with no network, watches its screen until Windows reaches its sign-in screen, keeps a screenshot as proof, then turns the VM off and deletes it with its files. It runs on a Proxmox cluster, using the same safeguards as Proxmox boot checks, or, for customers without Proxmox, on a Hyper-V host.
Setting it up #
- Agents → ⋮ on the machine → Restore points…, then Set up… beside Boot check.
- Choose Proxmox (cluster, node and storage for the copy) or Hyper-V (the host and a folder on it, for example
D:\Restores). - Set the copy's memory and processors, how long it may take to reach the sign-in screen (10 minutes unless you change it), and the days and time. The default is Saturdays at 13:00 UTC.
The newest result shows as a badge beside Boot check: Booted (green), Booted, no guest agent (amber: the copy was still running but its screen never settled on a sign-in screen), Boot check failed (red: it stopped, or stopped at a Windows blue screen) or Not checked with the reason. Click it for the screenshot, the restore point tested and the run log. A restore point that booted gets a tick in the list. Each check is also a run of the boot check job in the Jobs list, so Run now works there.
How it decides, and what it does to the host #
- The screen is the evidence. A machine imaged from hardware has no guest agent to ask, so the check waits for the screen to stop changing on something that is not blank: normally the sign-in screen. On Hyper-V, when Windows' own integration services answer, it also records the machine's name and Windows version.
- No network, ever. On Proxmox the copy gets a VM ID between 9100 and 9199 and its network card is read back as unplugged before it starts. On Hyper-V the VM's network adapter is not connected to any switch, read back from Hyper-V before it starts. There is no setting to turn this off.
- Only when there is room. The check is skipped, with the reason, when the host would have less than 4 GB of memory left beyond the copy's, or when the download folder or the storage would drop below a quarter free.
- Always removed. The copy is turned off and deleted, with its disks and any downloaded files, however the check ends. A copy left by an agent that was stopped mid-check is removed the next time a check runs (Hyper-V) or the agent starts (Proxmox). A copy that could not be removed sends an alert.
- What a pass means. Windows started from the backup and reached a steady screen. It does not test the applications inside.
The full screenshot is kept in the machine's own backup storage, next to its disk chains (machine-verify/ in the job's folder); Caryvane keeps only a small thumbnail. Failed checks are emailed like Proxmox boot checks; switch off Boot check failed in Settings → Notifications if you do not want them.
For single files, there is no need to restore the machine: download the VHDX and mount it in Windows, or keep a Files job alongside for everyday file restores.
Immutable storage and legal hold #
- Immutable backups. On storage with immutable backups on, each weekly base layer is locked for your lock period plus a week, like a Hyper-V full copy, so it outlives every nightly layer built on it.
- Legal hold. A legal hold covering the job stops old chains from being removed, whatever the version history says, until the hold is released.