Fossil

Backing Up a Remote Fossil Repository
Login

Backing Up a Remote Fossil Repository

One of the great benefits of Fossil and other distributed version control systems is that cloning a repository makes a backup. If you are running a project with multiple developers who share their work using a central server and the server hardware catches fire, the clones of the repository on each developer workstation may serve as a suitable backup.

We say “may” because it turns out not everything in a Fossil repository is copied when cloning. You don’t even always get copies of all historical file artifacts. More than that, a Fossil repository typically contains other useful information that is not always shared as part of a clone, which might need to be backed up separately. To wit:

Sensitive Information

Fossil purposefully does not clone certain sensitive information unless you’re logged in as a user with Setup capability. As an example, a local clone may have a different user table than the remote, because only a Setup user is allowed to see the full version for privacy and security reasons.

Configuration Drift

Fossil allows the local configuration to differ in several areas from that of the remote. You get a copy of some of these configuration areas on initial clone — not all! — but after that, remote configuration changes mostly do not sync down automatically.

Skin

Changes to the remote’s skin don’t sync down, on purpose, since you may want to have a different skin on the local clone than on the remote. You can ask for updates with fossil config pull skin, but that does not happen automatically during the course of normal development.

Email Alerts

The Admin → Notification settings do not get copied on clone or sync, and it is not possible to push such settings from one repository to another. We did this on purpose because you may have a network of peer repositories, and you only want one repository sending email alerts. If Fossil were to automatically replicate the email alert settings to a separate repository, subscribers would get multiple alerts for each event, which would be bad.

The only element of the email alert configuration that can be pulled over the sync protocol on demand is the subscriber list, via fossil config pull subscriber.

Project Configuration

This is normally generated once during fossil init and never changed, so Fossil doesn’t pull this information without being forced, on purpose. You could accidentally merge two separate Fossil repos by pushing one repo’s project config up to another, for example.

Others

A repo’s URL aliases, interwiki configuration, and ticket customizations also do not normally sync.

Private Branches

The very nature of Fossil’s private branch feature ensures that remote clones don’t get a copy of those branches. Normally this is exactly what you want, but in the case of making backups, you probably want to back up these branches as well. One of the two backup methods below provides this.

Shunned Artifacts

Fossil purposefully doesn’t sync shunned artifacts. If you want your local clone to be a precise match to the remote, it needs to track changes to the shun table as well.

Unversioned Artifacts

Data in Fossil’s unversioned artifacts table doesn’t sync down by default unless you specifically ask for it. Like local configuration data, it doesn’t get pulled as part of a normal fossil sync, but unlike the config data, you don’t get unversioned files as part of the initial clone unless you ask for it by passing the --unversioned/-u flag.

Autosync Is Intransitive

If you’re using Fossil in a truly distributed mode, rather than the simple central-and-clones model that is more common, there may be no single source of truth in the network because Fossil’s autosync feature isn’t transitive.

That is, if you cloned from server A, and then you stand that up on a server B, then if I clone from your server as my repository C, your changes to B autosync up to A, but not down to me on C until I do something locally that triggers autosync. The inverse is also true: if I commit something on C, it will autosync up to B, but A won’t get a copy until someone on B does something to trigger a sync there.

An easy way to run into this problem is to set up failover servers svr1 thru svr3.example.com, then set svr2 and svr3 up to sync with the first. If all of the users normally clone from svr1, their commits don’t get to svr2 and svr3 until something on one of the servers pushes or pulls the changes down to the next server in the sync chain.

Likewise, if svr1 falls over and all of the users re-point their local clones at svr2, then svr1 later reappears, svr1 is likely to remain a stale copy of the old version of the repository until someone causes it to sync with svr2 or svr3 to catch up again. And then if you originally designed the sync scheme to treat svr1 as the primary source of truth, those users still syncing with svr2 won’t have their commits pushed up to svr1 unless you’ve set up bidirectional sync, rather than have the two backup servers do pull only.

Solution 1: Explicit Pulls

The following script solves most of the above problems for the use case where you want a nearly-complete clone of the remote repository using nothing but the normal Fossil sync protocol. It only does so if you are logged into the remote as a user with Setup capability, however.

#!/bin/sh
fossil sync --unversioned
fossil configuration pull all
fossil rebuild

The last step is needed to ensure that shunned artifacts on the remote are removed from the local clone. The second step includes fossil conf pull shun, but until those artifacts are actually rebuilt out of existence, your backup will be “more than complete” in the sense that it will continue to have information that the remote says should not exist any more. That would be not so much a “backup” as an “archive,” which might not be what you want.

Solution 2: SQL-Level Backup

The first method doesn’t get you a copy of the remote’s private branches, on purpose. It may also miss other info on the remote, such as SQL-level customizations that the sync protocol can’t see. (Some ticket system customization schemes rely on this ability, for example.) You can solve such problems if you have access to the remote server, which allows you to get a SQL-level backup by delegating handling of locking and transaction isolation to the backup command, allowing the user to safely back up an in-use repository.

If you have SSH access to the remote server, something like this will work:

#!/bin/bash
bf=repo-$(date +%Y-%m-%d).fossil
ssh example.com "cd museum ; fossil backup -R repo.fossil backups/$bf" &&
    scp example.com:museum/backups/$bf ~/museum/backups

Beware that this method does not solve the intransitive sync problem, in and of itself: if you do a SQL-level backup of a stale repo DB, you have a stale backup! You should therefore run this on every node that may need to serve as a backup so that at least one of the backups is also up-to-date.

Encrypted Off-Site Backups

A useful refinement that you can apply to both methods above is encrypted off-site backups. You may wish to store backups of your repositories off-site on a service such as Dropbox, Google Drive, iCloud, or Microsoft OneDrive, where you don’t fully trust the service not to leak your information. This addition to the prior scripts will encrypt the resulting backup in such a way that the cloud copy is a useless blob of noise to anyone without the key:

iter=152830
pass="h8TixP6Mt6edJ3d6COaexiiFlvAM54auF2AjT7ZYYn"
gd="$HOME/Google Drive/Fossil Backups/$bf.xz.enc"
fossil sql -R ~/museum/backups/"$bf" .dump | xz -9 |
    openssl enc -e -aes-256-cbc -pbkdf2 -iter $iter -pass pass:"$pass" -out "$gd"

If you’re adding this to the first script above, remove the “-R repo-name” bit so you get a dump of the repository backing the current working directory.

Change the pass value to some other long random string, and change the iter value to something in the hundreds of thousands range. A good source for the first is here, and for the second, here.

You may find posts online written by people recommending millions of iterations for PBKDF2, but they’re generally talking about this in the context of memorizable passwords, where adding even one more character to the password is a significant burden. Given our script’s purely random maximum-length passphrase, there isn’t much more that increasing the key derivation iteration count can do for us.

Conversely, if you were to reduce the passphrase to 41 characters, that would drop the key strength by roughly 2⁶, being the entropy value per character for using most of printable ASCII in our passphrase. To make that lost strength up on the PBKDF2 end, you’d have to multiply your iterations by 2⁶ = 64 times. It’s easier to use a max-length passphrase in this situation than get crazy with key derivation iteration counts.

(This, by the way, is why the example passphrase above is 42 characters: with 6 bits of entropy per character, that gives you a key size of 252, as close as we can get to our chosen encryption algorithm’s 256-bit key size without going over. If it pleases you to give it 43 random characters for a passphrase in order to pick up those last four bits of security, you’re welcome to do so.)

Compressing the data before encrypting it removes redundancies that can make decryption easier, and it results in a smaller backup than you get with the previous script alone, at the expense of a lot of CPU time during the backup. You may wish to switch to a less space-efficient compression algorithm that takes less CPU power, such as lz4. Changing up the compression algorithm also provides some security-thru-obscurity, which is useless on its own, but it is a useful adjunct to strong encryption.

This requires OpenSSL 1.1 or higher. If you’re on 1.0 or older, you won’t have the -pbkdf2 and -iter options, and you may have to choose a different cipher algorithm; both changes are likely to weaken the encryption significantly, so you should install a newer version rather than work around the lack of these features.

Beware that macOS ships a fork of OpenSSL called LibreSSL that lacked this capability until Ventura (13.0). If you’re on Monterey (12) or older, we recommend use of the Homebrew OpenSSL package rather than give up on the security afforded by use of configurable-iteration PBKDF2. To avoid a conflict with the platform’s openssl binary, Homebrew’s installation is unlinked by default, so you have to give an explicit path to it, one of:

/usr/local/opt/openssl/bin/openssl ...     # Intel x86 Macs
/opt/homebrew/opt/openssl/bin/openssl ...  # ARM Macs (“Apple silicon”)

Restoring From An Encrypted Backup

The “restore” script for the above fragment is basically an inverse of it, but it’s worth showing it because there are some subtleties to take care of. If all variables defined in earlier scripts are available, then restoration is:

openssl enc -d -aes-256-cbc -pbkdf2 -iter $iter -pass pass:"$pass" -in "$gd" |
    xz -d | fossil sql --no-repository ~/museum/restored-repo.fossil

We changed the -e to -d on the openssl command to get decryption, and we changed the -out to -in so it reads from the encrypted backup file and writes the result to stdout.

The decompression step is trivial.

The last change is tricky: we used fossil sql above to ensure that we’re using the same version of SQLite to write the encrypted backup DB as was used to maintain the repository. We must also do that on restoration: Fossil serves as a dogfooding project for SQLite, often making use of the latest features, so it is quite likely that a given random sqlite3 binary in your PATH will be unable to understand the file created by “fossil sql .dump”! The tricky bit is, you can’t just pipe the decrypted SQL dump into fossil sql, because on startup, Fossil normally goes looking for tables created by fossil init, and it won’t find them in a newly-created repo DB. We get around this by passing the --no-repository flag, which suppresses this behavior. Doing it this way saves you from needing to go and build a matching version of sqlite3 just to restore the backup.