Restore RBD Image from Dead Ceph Cluster using only OSDs

Ceph is a great tool to store RBD images for VMs. It provides a distributed system with automated recovery. However, when everything fails it can be hard to extract the stored data. I recently faced this problem myself and after Googling for hours, trying to extract the important VM images from the OSDs, I finally managed to extract the Images.


Ceph configuration is lost, no monitors are working, internal recovery tools fail. But the OSDs are still present, i.e. there are no disk errors on the OSDs. The OSD service is not needed to extract the data.


Use these tools:

Ceph Data Recovery Tools

Tested with Ceph Hammer

What these tools are for

These tools are useful to rescue RBD Images stored on Ceph OSDs when everything except the OSDs are lost or broken. The tools allow you to extract the images from mounted OSDs without any Ceph Service running, so having physical OSDs mounted via network or locally is the only real requirement.


  • OSDs must be accessible (filesystem)
  • Have enough storage available to save the extracted images
  • Patience – some steps need much time
  • Screen? You should use screen when using ssh – some steps take multiple hours!


To recover the data, two steps are needed:

  1. Collect available data (Parts of the images distributed over all OSDs)
  2. Reassemble the blocks


  1. Clone this repo to a place of your choice. Make sure that you have at least a few hundrets MB space in this directory.
  2. Create a new subfolder osds in this folder.
  3. Choose one of the following options or mix them: 3.1. Attach all OSDs as local storage. For every OSD create a subfolder in the osds folder and mount it there. 3.2 Use sshfs to mount OSDs over network / ssh


Your directory structure should look like this:

| ..
| .
+- osds
 +- osd1
 +- osd2
 +- .......

Step 1: Collect files

Quite easy:
./ osds

This could take a bit longer. Depends on your mount strategy (local vs sshfs) and your network.

First result

You have some new folders now: vms and file_lists.
You should only be interested in the vms folder. It contains files named like your VM Images.
Use to print all VM Images found in step 1.

Step 2: Recover Image

Now we have everything to reassemble an image. The parts belonging to a specific Image are known and listed in files stored in file_lists.
To restore an image you need 3 information:

  1. The name of the Image (vms/
  2. The size of the original image in Bytes. So when the VM disk was 32GB in total (not used space!), you should use 34359738368.
    Important: If you are unsure about the actual disk size, choose a size which is larger! You can add some Bytes, MBytes or GBytes just to be sure
  3. A destination folder. Just a folder with enough free space to store your image of the specified size. (e.g. /mnt/sda)

Having these 3 information you can restore the image:
./ vms/ 34359738368 /mnt/sda

This will process all parts of the image and write it to a single image file. After this you can mount this image and access data or just put it back to a new cluster.

Repeat this for every disk image you need.

4 Comments on “Restore RBD Image from Dead Ceph Cluster using only OSDs”

  1. Hello,

    I’m trying to use Your tool to extract data from broken ceph cluster. It looks like Your tool figured out proper volumes ID, but is finishing very fast for 500 GB images. Can You advice, what to do to make it work ? Our ceph is on jewel, not hammer, I think it might be a problem. Please respond, I have huge problem with this broken cluster and Your tool is last chance to recover this data.


    • Hi,
      how many parts does assemble detect for you? There should be some output like

      “There are $count blocks found”.

      Could you send me the full output of the script?

      However, as far as I know the naming of the different parts did not change from hammer to jewel, but I am not sure.

      How many OSDs do you have and are all of them mounted on your re-assemble machine?


    • I actually don’t know and have not looked into this.

      You can still try to follow the procedure described here and see if the detection still works. If yes, try to recover a small Disk Image for testing purposes.

      Since the FileStore and BlueStore are compatible I would assume that there is a good chance that it still works.
      Please let me know if your attempts were successful!

Leave a Reply

Your email address will not be published. Required fields are marked *