It would be a good idea to document what would be the increased risk of failure resulting from selecting a subset of available blockdevs to write metadata to

[mulkieran]

[agrover]

How do you mean? If we have a 1000 disk pool for example and write metadata to 10, if those 10 somehow end up not being in the pool after a reboot, then we will get stale metadata and that could be extremely bad.

[mulkieran]

Yes. That's one failure mode. But we should understand it in terms of probabilities, like people are able to do with RAID.