Perhaps io_uring is a way forward. Today you can already batch a bunch of IO calls and chain them together so that the next is only excuted when the previous on

Perhaps io_uring is a way forward. Today you can already batch a bunch of IO calls and chain them together so that the next is only excuted when the previous one succeeds. With some additional hints journaling and similar filesystems could try putting such a batch into a separate transaction group. That would give you some crash all-or-nothing behavior but no rollback on error. It would also be on a best-effort basis, might only support some systemcalls etc. Then the situation could be improved piece by piece.