How to properly achieve asymmetric multi-site backups

[Tectu]

Let there be three hosts: A, B and C. All of them use ZFS.

• A is the primary production server.
• B is an on-site backup destination host with plenty of storage.
• C is a remote backup destination host with not so much storage.

What I'd like to achieve is to backup everything from host A to host B. This is what is working nicely so far.
On top of that, I'd like to have host B backup some datasets to host C.
What is the proper, correct way of achieving this?

On host A I use the ZFS property syncoid:sync to select which datasets to push to host B (or rather, which datasets not to push).

I basically see two options to achieve this:

1. On host B, also use the syncoid:sync property to select which datasets to push to host C.
2. On host A, set the syncoid:sync property to either B or B,C selectively.

Specifically I'd like to ask:

1. Are the two options I listed above correct / reasonable?
2. Are there any other options worth considering?
3. Which of the two options makes more sense? Personally, I feel like option 1. is better for my particular scenario/taste as I don't have to mess with the hostname resolution briefly mentioned in the sanoid readme.

[jimsalterjrs]

Personally, I just put things in a hierarchical order that works for me, and use recursion.

In other words, put the stuff to back up to C below the stuff to back up to B. So, you have a data structure on host A like poolA/thingsforB/thingsforC. Then, root@B:~# syncoid -r root@A:poolA/thingsforB poolB/thingsforB, followed by root@C:~# syncoid -r root@B:poolB/thingsforB/thingsforC poolC/thingsforC.

If for some reason you can't possibly organize your datasets hierarchically, well, you already found the sync property. :)

The biggest thing that usually trips people up about n-way replication is that the sync snapshots don't really work well anymore, because they get replicated to foreign machines that don't know what to do with them (and therefore they pile up and run you out of disk space, if you don't manually keep up with them).

Generally, if you're doing A->B and A->C (or any other topology of more-than-two-way replication) you should be using --no-sync-snap to keep syncoid from creating its own ephemeral snapshots, and just rely on sanoid on the source machine to provide regular snapshots.

Obviously, you do need to monitor your ongoing replication to make sure you don't fall out of sync--otherwise, if you fail to replicate too many times in a row, you eventually wind up with no common snapshots with the source.

[Tectu]

Thank you for the quick answer on this!

The hierarchical approach does not work for some of my situations.

Generally, if you're doing A->B and A->C [...]

The thing I actually want to do is A -> B and B -> C. I do not want A to know anything about C.
Just with the added notion that I want to selectively pick which datasets that A pushed to B are pushed to C. In other words, C is only a "partial copy" of B.

I'm already making use of --no-sync-snap.

Is my understanding correct that therefore I should go with option 1 outlined in my opening post?
Is that a "common" approach?