Backing up a mail server

[upnort]

How does the new Proxmox Backup Server (PBS) scale?

I have one specific use case in mind: a CentOS 7 LXC container mail server running Postfix/Dovecot . There is about 300 GB of data at about 3.5 million mostly small files. The volume is continually increasing.

I am aware of Dovecot replication and I'm looking into that. Possibly replication is the only viable strategy, but I'm curious how PBS might handle this kind of volume.

Ideal would be some kind of incremental or differential vzdump. Mail servers are horribly dynamic. The longer the interval between backups the more mail data that is lost in a disaster. The bottom line is in the event of a host failure that the LXC container can be restored to a new host with as little loss of files as practical.

Thank you for any insight.

 

[t.lamprecht]

Hi,

that workload is probably rather on the slower end of the spectrum in terms of duration required for assembling the backup, the resulting backup on the other hand should be quite good regarding storage space efficiency and also restore performance.

A few architectural notes here:
* Proxmox Backup Server always wants to do a full backup, i.e., a self contained state to avoid the issues with applying deltas. It still only sends the delta of chunks, but the resulting archive is fully self contained.
* PBS works with a sliding window algorithm for file level backup, this batches more small files together efficiently and speeds up handling the backup (pruning, garbage collection, ...) on the server side, avoiding the need for full locking like other approaches (IIRC, e.g., borg) do. This results in the need to read the whole directory tree, which has naturally it's impact.
But, it should be still faster as a plain old vzdump was, and naturally other perks like the deduplication help on PBS storage side.

If you'd use VMs this would be not a big issue, as there we work on the blocklevel and the dirty-bitmap can tell us which blocks changed in an instant, which makes it really fast there.

IMO, a application aware replication can be much more efficient that a "bigger hammer" like backing up the whole underlying FS very often.
That's also why database replication is almost never replaced by replicating the underlying storage but with builtin technology.

 

[upnort]

Thank you!