For some reason VM and DB workloads are btrfs's Achilles heel but ZFS seems to handle them pretty well (provided that a suitable recordsize is set). How do they perform on bcachefs?
The issue is that also disables many of the interesting features of BTRFS for those files. No checksumming, no snapshots and no compression. In comparison ZFS handles these features just fine for those kinds of files without the enormous performance / fragmentation issues of BTRFS (without nodatacow).
chattr +C does not disable snapshots. It switches the writes from CoW-everything to CoW-only-after-snapshot. Snapshots work just fine, with only the first write (per extent) after a snapshot paying the CoW cost.
