PBS dimensioning

[Kuonel]

Hi everyone, it's my first post, nice to meet you all

I'm running an 8 nodes proxmox cluster with hyperconverged ceph on an SFP+ network,
all nvme pci 4, two 4tb drives for each node, 58tb in total (as reported by ceph).

I'm currently using 16tb of data
and I'd like to add a PBS box.

I've read that an all SSD solution would be ideal
and that two drives must be dedicated to the OS.

My day-to-day storage usage doesn't grow much
but I write and delete lots of data (let's say 1tb each day)
and so I'm having trouble dimensioning the drives for PBS.

I'm inclined towards 8 x Samsung MZ-77Q8T0BW (8tb cheap sata ssd)
and two small sata ssd for the OS.

I would use the same SFP+ network as ceph, limiting the bandwidth usage for the backups to around 300MB/s.


Questions:

1.
Am I correct to assume that a daily backup would grow the snapshots by 1tb each day ?

2.
Does PBS provide a 'rotation policy', which deletes the old snapshots after X days ?

3.
Do I really need an all SSD solution
or can I use spinning rust + a couple of sata ssd for caching
given that I have all night (8 hours) for the backups ?



Thank you,
Mike

 

[Dunuin]

Does PBS provide a 'rotation policy', which deletes the old snapshots after X days ?

You can set a backup retention: https://pbs.proxmox.com/docs/prune-simulator/

Am I correct to assume that a daily backup would grow the snapshots by 1tb each day ?

If it is new data that isn't well compressible and not deduplicatable, yes.

Do I really need an all SSD solution
or can I use spinning rust + a couple of sata ssd for caching
given that I have all night (8 hours) for the backups ?

PBS needs IOPS performance. 32TB of backups means roughly 16-32 million chunk files (1-2 MB each, max 4MB). GC should be OK with SSD for metadata but a verify task would take really long.

Also keep in mind that consumer SSDs are terrible with ZFS and QLC SSDs like those Samsung disks are exceptional horrible...

 

[Kuonel]

Thank you Dunuin for your inputs