While there's probably a better way of doing it via the docker zfs driver, I just make a datastore per stack under the hypervisor, mount the datastore into the docker LXC, and make everything bind mount within that mountpoint, then snapshot and backup via Sanoid to a couple of remote ZFS pools, one local and one on zfs.rent.
I've had to restore our mailserver (mysql) and nextcloud (postgres) and they both act as if the power went out, recovering via their own journaling systems. I've not found any inconsistencies on recovery, even when I've done a test restore on a snapshot that's been backed up during known hard activitiy. I trust both databases for their recovery methods, others maybe not so much. But test that for yourself.