It's a good question.

The benefit of variant here is that, while we will enforce schemas in silver, there will be many use-cases where users are looking to query columns that we do not have in silver. The reason is that new columns are added by users somewhat on the fly. They create a configuration file that collects sensor measurements from a variety of available sensors and the sensor data are what come through autoloader. If they have a new reading that they added to the config that we haven't explicitly unpacked in a silver table, they will need to query bronze directly, in which case variant is much more user-friendly, especially since these users will not be experienced in Databricks. Conversely, with an ever-expanding wide table format, the vast majority of columns will be NULL for a given row which will make the user experience more confusing.