I see, I was not precise in my wording: I'm running elasticsearch/kibana in a k8s cluster (elasticsearch running as a statefulset). When I say "simply update", I mean that I just replace the pointer to the docker images in that statefulset, the effect is that the running containers are replaced and the data directories are preserved (that's the rolling upgrade as referenced by you).
As said, this one works fo me.
What doesn't work: I have regular backups (elasticsearch snapshot API) stashed into a gcs bucket. When I restore a snapshot that was created from 6.7 directly into an empty 7.0 elasticsearch/kibana installation, then it fails with the error discussed in this thread.