At the moment I am using the latter approach with bind volumes in a Container going to a host mounted network share, which is then backed up or a regular basis, runs on RAID etc.
The one downside I see in this is that if there is a catastrophic failure, there won’t be automatic fail-over to some other replica and any restore will only be as good as the last backup.
Be interested to see what other folks are doing.
Edit: Thinking about this, if something takes out the entire storage array, then Kafka will be the least of our problems!