Find out why dataset height rolls back

[kalabukdima]

Currently, the last dataset block known (from metadata written by scheduler) by the portal sometimes decreases.
It can't be the case that parquet chuks are removed from S3. We should find out whether it's a bug on the scheduler side or in the portal.

[kalabukdima]

It gets written like that by the scheduler indeed. Probably because it restarts and loses state sometimes.
It will be fixed when we have a new scheduler.