If Scrubs Hurt, Your ZFS Design Is Broken

I have attended some of the webinars and listen to Allan on a podcast regularly and this LLM-written post is painful to read. If I hadn't already been familiar with him and his work I would have not given much weight to what was written. He is a big proponent of tuning ZFS which requires knowing your workload for which he has written other guides.

1. From my own experience, most people who use a ZIL for faster sync writes are wasting space. You only need five seconds of peak throughput (for a 10Gb connection that's ~5 GB). Create a pair of 25 GB partitions for the ZIL rather than dedicating entire devices.

2. With the remaining space set up a mirrored Special VDEV with heavy use of small blocks. 32k block sizes will still give some room for compression which, other than saving space, can increase throughput and reduce IO/ps.

3. Pause scrubs (`zpool scrub -p`) when the workload is either expected to be heavy or is actually heavy based on moving averages and resume when its safe. Even my 24 TB NAS takes 24+ hours with no other activity.

> Sizing is measurable rather than guesswork. zdb -bb tank breaks down allocated space by block type; sum the non-data categories and you have your current metadata footprint.

Of course, this is also not a fast command on a slow array; my half full 2x 24TB mirror is estimating ~ 12 hours to run this; the last scrub took just over 24 hours to run. (After I posted, the remaining time dropped to 10 minutes... not as slow as it seemed, but still slow, and using a lot of I/O ... left for lunch and 50 minutes later it's almost done; estimation is hard)

Every 100+ TB ZFS pool I have is optimized for streaming reads/writes. IOPS are rarely a binding factor at these sizes.

ZFS scrub performance shouldn’t come at the expense of production worklaods. If scrubs cause latency or downtime, your pool lacks the IOPs and metadata performance needed for healthy maintanance.

This could have been a decent article if it hadn't been run through the AI slop machine. If I wanted to read writing this bad I would have just asked ChstGPT myself.

It's so hard to wade through run on sentences like "The knobs are not the problem. The pool was designed without room for this workload."

Oh sure, just buy more disks. Who do you think I am, that guy on the Monopoly box?

I stocked up on a few shuckable externals before the start of the second Grump catastrophe. $280/24TB (Exos inside). I'm finally getting around to setting up a pool, thinking 4+1 is kind of poor redundancy and wondering about adding another disk. The things are up to over $1,000 each, in the husk or not. Fucking hell. It makes me want to drastically scale back the size of my new pool in case a drive does die so I don't end up needing to buy a replacement any time soon.