Btrfs has separate metadata and data block groups, allocated dynamically since some workloads can be light or heavy on metadata. And the data:metadata usage rat

Btrfs has separate metadata and data block groups, allocated dynamically since some workloads can be light or heavy on metadata. And the data:metadata usage ratio can change after allocation and freeing space by deleting files. The idea of converting between block groups is being kicked around. At this point it is kind of a "last mile" sort of problem, not that common. But annoying for those who run into it because it's non-obvious how to get out of it. Safest option is to just look at 'btrfs fi us' and see which type of block group is the least full. Do a filtered balance on it. e.g. Use something like '-dusage=5' to essentially get rid of any data block groups that are 5% or less full. Their data extents get moved to other block groups. Now that space is "unallocated" and can be allocated into new metadata block groups. These sorts of gymnastics should be rare. There is https://github.com/kdave/btrfsmaintenance https://github.com/kdave/btrfsmaintenance that will do them automatically for those who encounter the problem with their particular workload more often than they care to deal with manually. I never have to manually balance, and I also don't use the automated scripts - it's been years since I've hit enospc that wasn't a case of a 100% full file system.