Are azure disk snapshots not atomic?

[scoopex]

One of my backup mechanisms of my postgres databases running on a AKS cluster data is based on velero disk snapshots.

I implemented a automated test procedure to verify my active backup-mechanisms. This procedure automatically runs a backup every night at 2AM, deletes a namespace and restores that namespace at 3AM (a situation of very low activity).

Unfortunately the restored postgresql database (pvc/pv created for the snapshot) does not start properly after restart with the following messages.

2021-04-19 10:42:55.008 GMT [179] LOG:  database system was interrupted; last known up at 2021-04-18 02:00:49 GMT
2021-04-19 10:42:55.096 GMT [179] LOG:  invalid primary checkpoint record
2021-04-19 10:42:55.096 GMT [179] PANIC:  could not locate a valid checkpoint record
2021-04-19 10:42:55.204 GMT [1] LOG:  startup process (PID 179) was terminated by signal 6: Aborted
2021-04-19 10:42:55.204 GMT [1] LOG:  aborting startup due to startup process failure
2021-04-19 10:42:55.210 GMT [1] LOG:  database system is shut down

In my understanding a postgresql-situation like this only can happen if disk snapshots are not atomic.
Obviously, a restored snapshot is always in a unclean mount state and unclean database state.

The data located on the snapshot should be consistent in the moment of the snapshot from a point of view of binary write operations order. I my situation this not seems to be true because the database is not capable to recover in a automated way after the filesystem recovery.

Are azure disk snapshots not atomic and how do disk-snapshots make sense if they are not?