← sources

ZFS and cheap SSDs - using a "sacrificial" SLOG disk to protect other SSDs from wearout, would it work?

source: reddit (reddit:homelab)published 2026-10-29 05:46ingested 2026-10-30 02:39

retained text

So with AI driving storage prices sky-high, even more so for a country way down the bottom of the Pacific, I'm thinking of a novel approach to not having to replace worn SSDs in my homelab/homebizlab so often...

**The setup**

I have a three-node Proxmox setup built out of some ex-lease workstations of varying ages.

I am using consumer-level SSDs (Samsung 970 EVOs in one node, 990 PROs in another) because enterprise-level is basically unaffordable.

It's set up as HA, with storage being Linstor over ZFS. I've chosen this because it allows node 3 to access storage that's replicated only across nodes 1+2, and it performs much better than Ceph at this small scale, which I had previously and was just too slow.

The ZFS arrays are:

* `rpool` Proxmox default on each node - a ZFS mirrored pool on the first 40GB of 2x 1TB SSDs on each node
* `local-tier1` \- ZFS pool, local to each of nodes 1 and 2 (but not 3) - on the next 400GB of the 2x 1TB SSDs on each node. The remaining capacity of these 1TB SSDs is not provisioned.
* `drbd-tier2` layered on top of `local-tier1` and replicated between nodes 1 and 2, accessible from node 3
* `local-tier3` \- ZFS pools local to each of nodes 1 and 2, on mirrored HDDs (6TB)

**The problem**

The main issue with this setup is that Proxmox hammers the SSDs due to corosync's continuous synchronous writes to rpool. Wearout increases rapidly, my previous 512GB Samsung 980s (which the 990 PROs just replaced) took just 2.5 years to hit 100% wearout and have SMART give pre-failure warnings. They're right on warranty with their TBW.

They cost 8x as much to replace (1TB as 512GB unavailable) as they had to originally buy.

On the `rpool` pool, `sync=standard` is set as I think is default. I could set this to `none` but then I risk data loss or messy recovery after a crash etc as corosync's synchronous writes only get written every `zfs_txg_timeout` seconds (I'm using the default, 5).

But this morning I thought of an alternative, and want to validate this here to see if this works the way I think it would?

**My either-genius-or-stupid idea - SLOG sacrificial disks to take the sync write wear**

The way I understand it, a SLOG disk normally operates write-only. When `sync=standard` and a SLOG disk is present, any synchronous writes are immediately written to the SLOG disk (assuming that disk honours the , then later get written asynchronously to the main pool either when `zfs_txg_timeout` is reached or 64MB is ready to write out.

Again if I understand it right, the SLOG is only read if coming back up after a dirty shutdown, to flush its contents to the pool.

Which tells me that when doing lots of synchronous writes, the SLOG disk will take the brunt of the wear, and not so much the main SSDs?

So I'm thinking if I add a single "sacrificial" SSD as a SLOG for the `rpool` and `local-tier1` pools, *not* to increase performance, but to protect the main SSDs from excessive wear from all the sync writes - would it work as I am thinking?

The sacrificial SSD could be one I don't fully trust - a cheapie, one lying around too small for anything else, or one with already high wearout. I have a small pile of these in a cupboard.

My understanding is if the SLOG disk fails, because the ZIL is not written *through*, unless it's unfortunate enough to fail in the same moment the host dies for any other reason, the ZIL still exists in RAM and then the earlier behaviour (synchronous writes directly to the ZIL on the disks holding the `rpool` pool) would simply resume until I can swap it out with the next sacrificial one.

Have I got that right? Would it work as I am thinking?

edit 4h later: corrected all references to "ZLOG" to "SLOG" - oops  
also see update comment - tried it on one node, initial signs say it works!

document fetched 2026-10-30 00:39 · canonical https://reddit.com/r/homelab/comments/1xibpz0/zfs_and_cheap_ssds_using_a_sacrificial_slog_disk