Additional Notes

Here are misc. notes about topics that are maybe not covered in enough detail in the usage section.


The chunker params influence how input files are cut into pieces (chunks) which are then considered for deduplication. They also have a big impact on resource usage (RAM and disk space) as the amount of resources needed is (also) determined by the total amount of chunks in the repository (see Indexes / Caches memory usage for details).

--chunker-params=buzhash,10,23,16,4095 results in a fine-grained deduplication and creates a big amount of chunks and thus uses a lot of resources to manage them. This is good for relatively small data volumes and if the machine has a good amount of free RAM and disk space.

--chunker-params=buzhash,19,23,21,4095 (default) results in a coarse-grained deduplication and creates a much smaller amount of chunks and thus uses less resources. This is good for relatively big data volumes and if the machine has a relatively low amount of free RAM and disk space.

--chunker-params=fixed,4194304 results in fixed 4MiB sized block deduplication and is more efficient than the previous example when used for for block devices (like disks, partitions, LVM LVs) or raw disk image files.

--chunker-params=fixed,4096,512 results in fixed 4kiB sized blocks, but the first header block will only be 512B long. This might be useful to dedup files with 1 header + N fixed size data blocks. Be careful to not produce a too big amount of chunks (like using small block size for huge files).

If you already have made some archives in a repository and you then change chunker params, this of course impacts deduplication as the chunks will be cut differently.

In the worst case (all files are big and were touched in between backups), this will store all content into the repository again.

Usually, it is not that bad though:

  • usually most files are not touched, so it will just re-use the old chunks it already has in the repo

  • files smaller than the (both old and new) minimum chunksize result in only one chunk anyway, so the resulting chunks are same and deduplication will apply

If you switch chunker params to save resources for an existing repo that already has some backup archives, you will see an increasing effect over time, when more and more files have been touched and stored again using the bigger chunksize and all references to the smaller older chunks have been removed (by deleting / pruning archives).

If you want to see an immediate big effect on resource usage, you better start a new repository when changing chunker params.

For more details, see Chunks.

--noatime / --noctime

You can use these borg create options to not store the respective timestamp into the archive, in case you do not really need it.

Besides saving a little space for the not archived timestamp, it might also affect metadata stream deduplication: if only this timestamp changes between backups and is stored into the metadata stream, the metadata stream chunks won’t deduplicate just because of that.

--nobsdflags / --noflags

You can use this to not query and store (or not extract and set) flags - in case you don’t need them or if they are broken somehow for your fs.

On Linux, dealing with the flags needs some additional syscalls. Especially when dealing with lots of small files, this causes a noticeable overhead, so you can use this option also for speeding up operations.


borg uses a safe default umask of 077 (that means the files borg creates have only permissions for owner, but no permissions for group and others) - so there should rarely be a need to change the default behaviour.

This option only affects the process to which it is given. Thus, when you run borg in client/server mode and you want to change the behaviour on the server side, you need to use borg serve --umask=XXX ... as a ssh forced command in authorized_keys. The --umask value given on the client side is not transferred to the server side.

Also, if you choose to use the --umask option, always be consistent and use the same umask value so you do not create a mixup of permissions in a borg repository or with other files borg creates.


The --read-special option is special - you do not want to use it for normal full-filesystem backups, but rather after carefully picking some targets for it.

The option --read-special triggers special treatment for block and char device files as well as FIFOs. Instead of storing them as such a device (or FIFO), they will get opened, their content will be read and in the backup archive they will show up like a regular file.

Symlinks will also get special treatment if (and only if) they point to such a special file: instead of storing them as a symlink, the target special file will get processed as described above.

One intended use case of this is backing up the contents of one or multiple block devices, like e.g. LVM snapshots or inactive LVs or disk partitions.

You need to be careful about what you include when using --read-special, e.g. if you include /dev/zero, your backup will never terminate.

Restoring such files’ content is currently only supported one at a time via --stdout option (and you have to redirect stdout to where ever it shall go, maybe directly into an existing device file of your choice or indirectly via dd).

To some extent, mounting a backup archive with the backups of special files via borg mount and then loop-mounting the image files from inside the mount point will work. If you plan to access a lot of data in there, it likely will scale and perform better if you do not work via the FUSE mount.


Imagine you have made some snapshots of logical volumes (LVs) you want to backup.


For some scenarios, this is a good method to get “crash-like” consistency (I call it crash-like because it is the same as you would get if you just hit the reset button or your machine would abruptly and completely crash). This is better than no consistency at all and a good method for some use cases, but likely not good enough if you have databases running.

Then you create a backup archive of all these snapshots. The backup process will see a “frozen” state of the logical volumes, while the processes working in the original volumes continue changing the data stored there.

You also add the output of lvdisplay to your backup, so you can see the LV sizes in case you ever need to recreate and restore them.

After the backup has completed, you remove the snapshots again.

$ # create snapshots here
$ lvdisplay > lvdisplay.txt
$ borg create --read-special /path/to/repo::arch lvdisplay.txt /dev/vg0/*-snapshot
$ # remove snapshots here

Now, let’s see how to restore some LVs from such a backup.

$ borg extract /path/to/repo::arch lvdisplay.txt
$ # create empty LVs with correct sizes here (look into lvdisplay.txt).
$ # we assume that you created an empty root and home LV and overwrite it now:
$ borg extract --stdout /path/to/repo::arch dev/vg0/root-snapshot > /dev/vg0/root
$ borg extract --stdout /path/to/repo::arch dev/vg0/home-snapshot > /dev/vg0/home

Separate compaction

Borg does not auto-compact the segment files in the repository at commit time (at the end of each repository-writing command) any more.

This is new since borg 1.2.0 and requires borg >= 1.2.0 on client and server.

This causes a similar behaviour of the repository as if it was in append-only mode (see below) most of the time (until borg compact is invoked or an old client triggers auto-compaction).

This has some notable consequences:

  • repository space is not freed immediately when deleting / pruning archives

  • commands finish quicker

  • repository is more robust and might be easier to recover after damages (as it contains data in a more sequential manner, historic manifests, multiple commits - until you run borg compact)

  • user can choose when to run compaction (it should be done regularly, but not necessarily after each single borg command)

  • user can choose from where to invoke borg compact to do the compaction (from client or from server, it does not need a key)

  • less repo sync data traffic in case you create a copy of your repository by using a sync tool (like rsync, rclone, …)

You can manually run compaction by invoking the borg compact command.

Append-only mode (forbid compaction)

A repository can be made “append-only”, which means that Borg will never overwrite or delete committed data (append-only refers to the segment files, but borg will also reject to delete the repository completely).

If borg compact command is used on a repo in append-only mode, there will be no warning or error, but no compaction will happen.

append-only is useful for scenarios where a backup client machine backups remotely to a backup server using borg serve, since a hacked client machine cannot delete backups on the server permanently.

To activate append-only mode, set append_only to 1 in the repository config:

borg config /path/to/repo append_only 1

Note that you can go back-and-forth between normal and append-only operation with borg config; it’s not a “one way trip.”

In append-only mode Borg will create a transaction log in the transactions file, where each line is a transaction and a UTC timestamp.

In addition, borg serve can act as if a repository is in append-only mode with its option --append-only. This can be very useful for fine-tuning access control in .ssh/authorized_keys:

command="borg serve --append-only ..." ssh-rsa <key used for not-always-trustable backup clients>
command="borg serve ..." ssh-rsa <key used for backup management>

Running borg init via a borg serve --append-only server will not create an append-only repository. Running borg init --append-only creates an append-only repository regardless of server settings.


Suppose an attacker remotely deleted all backups, but your repository was in append-only mode. A transaction log in this situation might look like this:

transaction 1, UTC time 2016-03-31T15:53:27.383532
transaction 5, UTC time 2016-03-31T15:53:52.588922
transaction 11, UTC time 2016-03-31T15:54:23.887256
transaction 12, UTC time 2016-03-31T15:55:54.022540
transaction 13, UTC time 2016-03-31T15:55:55.472564

From your security logs you conclude the attacker gained access at 15:54:00 and all the backups where deleted or replaced by compromised backups. From the log you know that transactions 11 and later are compromised. Note that the transaction ID is the name of the last file in the transaction. For example, transaction 11 spans files 6 to 11.

In a real attack you’ll likely want to keep the compromised repository intact to analyze what the attacker tried to achieve. It’s also a good idea to make this copy just in case something goes wrong during the recovery. Since recovery is done by deleting some files, a hard link copy (cp -al) is sufficient.

The first step to reset the repository to transaction 5, the last uncompromised transaction, is to remove the hints.N, index.N and integrity.N files in the repository (these files are always expendable). In this example N is 13.

Then remove or move all segment files from the segment directories in data/ starting with file 6:

rm data/**/{6..13}

That’s all to do in the repository.

If you want to access this rolled back repository from a client that already has a cache for this repository, the cache will reflect a newer repository state than what you actually have in the repository now, after the rollback.

Thus, you need to clear the cache:

borg delete --cache-only repo

The cache will get rebuilt automatically. Depending on repo size and archive count, it may take a while.

You also will need to remove ~/.config/borg/security/REPOID/manifest-timestamp.


As data is only appended, and nothing removed, commands like prune or delete won’t free disk space, they merely tag data as deleted in a new transaction.

Be aware that as soon as you write to the repo in non-append-only mode (e.g. prune, delete or create archives from an admin machine), it will remove the deleted objects permanently (including the ones that were already marked as deleted, but not removed, in append-only mode). Automated edits to the repository (such as a cron job running borg prune) will render append-only mode moot if data is deleted.

Even if an archive appears to be available, it is possible an attacker could delete just a few chunks from an archive and silently corrupt its data. While in append-only mode, this is reversible, but borg check should be run before a writing/pruning operation on an append-only repository to catch accidental or malicious corruption:

# run without append-only mode
borg check --verify-data repo && borg compact repo

Aside from checking repository & archive integrity you may want to also manually check backups to ensure their content seems correct.

Further considerations

Append-only mode is not respected by tools other than Borg. rm still works on the repository. Make sure that backup client machines only get to access the repository via borg serve.

Ensure that no remote access is possible if the repository is temporarily set to normal mode for e.g. regular pruning.

Further protections can be implemented, but are outside of Borg’s scope. For example, file system snapshots or wrapping borg serve to set special permissions or ACLs on new data files.

SSH batch mode

When running Borg using an automated script, ssh might still ask for a password, even if there is an SSH key for the target server. Use this to make scripts more robust:

export BORG_RSH='ssh -oBatchMode=yes'