A need for a way to manually mark a device as "DEGRADED" (or "SUSPECT") (zpool degrade $DISK ?)

[ptrrkssn]

I've been burned by this multiple times now... The situation:

1. You have a RAIDZ2 pool.
2. You have a drive that is starting to go bad. SMART reports " Firmware impending failure data error rate too high" but it still works and doesn't generate any errors that ZFS can detect (but might be starting to be slower).
3. You want to be a good sysadmin and proactively replace it so does a: "zpool replace POOL BAD NEW" where NEW is one of the hotspares, but don't offline the drive since it still works so why reduce the redundancy when you don't have too.
4. You wait for resilver to complete
5. Then you find that FreeBSDs zfsd (daemon that monitors disks and automatically activates hotspares and things) have noticed that the resilver is complete, and finds two ONLINE disks in that zvol -and so de it detaches the hotspare again, and you're back at the same situation as before - just 5 days later.
6. You sigh and go and have a coffee... and then start the whole process again.

(Yes, I should have removed NEW from the hotspares "pool" before starting the replace operation, but I forgot, that probably would have stopped zfsd from interfering)

Having a way to manually mark drives as DEGRADED (or SUSPECT or something) telling ZFS (and zfsd) to avoid reactivating when there is redundancy, but still be able to use it as a last resort would have been nice :-)

[lundman]

Will zpool offline pool device not do?

[ptrrkssn]

Sure. But then ZFS stops using the disk completely. I was trying for some way to just hint ZFS to just use it as a "last resort" since it still works (but probably is a little bit slower due to have to seek more or retry more).

[tonyhutter]

Partial solution:

zpool set allocating=off <pool> <bad_vdev>

[tonyhutter]

Also consider:

zpool set sit_out=on <pool> <bad_vdev>

This will force ZFS to reconstruct read data from parity, rather than from reading bad_vdev directly (although it will fallback to reading from bad_vdev if it can't reconstruct)